Communication method and apparatus

Abstract

The present application provides a communication method and an apparatus, and relates to the technical field of communications. In the method, a second network element can send first information to other network elements to notify a time of data used for model inference, so that each network element participating in distributed inference can determine inference data on the basis of the time. Negotiation of a time for inference data enables each network element participating in distributed inference to determine data having a closer time, so as to use same for model inference, thus improving the accuracy of determining inference results.

Description

Communication method and device

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to the Chinese patent application filed with the State Intellectual Property Office of the People's Republic of China on March 19, 2024, with application number 202410319900.6 and application name "A Communication Method and Device", the entire contents of which are incorporated by reference into this application. Technical Field

[0003] The present application relates to the field of communication technology, and in particular to a communication method and device. Background Art

[0004] Distributed learning (or distributed training) refers to a learning method that combines data from multiple participants to jointly train a model. During the model training phase, the initiator receives a task to be processed and determines the data required by the distributed learning model to handle it. Since the initiator may not be able to provide all the data required to train the distributed learning model, it must collaborate with other participants to train the model. During the distributed learning process, each participant obtains its own training data and trains its local model based on that data. Participants other than the initiator can send training results to the initiator, which then determines the global (or overall) training loss based on these and the initiator's training results. This effectively trains the model based on the training data of each participant, but participants do not need to exchange their training data to ensure data security. After distributed learning, each participant obtains its own trained model. Accordingly, during the model inference phase, the initiator obtains the inference results of other participants and, based on these and the initiator's own results, derives the global inference result.

[0005] Currently, in the model reasoning stage, each participant performs model reasoning based on their own reasoning data, which makes it difficult to ensure the accuracy of the reasoning results. Summary of the Invention

[0006] The embodiments of the present application provide a communication method and apparatus for improving the accuracy of model reasoning.

[0007] In a first aspect, an embodiment of the present application provides a communication method. The method can be applied to a communication system comprising a first network element and a second network element. In an embodiment of the present application, a network element (such as any one of the first network element and the second network element) can be a device, such as a core network device, an access network device, a terminal device, or a server. Alternatively, the network element can also be a chip system (or, chip) or other functional modules. The other functional modules include, for example, one or more software modules (such as computer programs), hardware modules, or hardware modules running programs, and are not specifically limited to this.

[0008] The method includes: a first network element receiving first information from a second network element, the first information indicating a time for determining inference data, the inference data being used for inference based on a first model, the first model being associated with one or more models; the first network element determining an inference result of the first model based on the first data and the first model, the first data being inference data determined based on the first information; the first network element sending second information to the second network element, the second information indicating an inference result of the first model; and the second network element determining the first inference result based on the second information and third information, the third information indicating an inference result of a second model, the second model being one of the one or more models. For example, the second network element may aggregate the inference result of the first model with the inference result of the second model to obtain the first inference result. Each network element participating in model inference may be a participant, and a network element coordinating the model inference may be a coordinator. The coordinator may not be considered a participant in model inference. The party initiating model inference may be a participant or a coordinator, without specific limitation.

[0009] The second network element may be a participant in model inference. In this case, the one or more models include a model corresponding to the first network element (e.g., the first model) and a model corresponding to the second network element. For example, the model corresponding to the second network element may be the second model. The first model may replace the model described as the first network element, for example, the local model of the first network element. Similarly, the second model may replace the model described as the second network element, for example, the local model of the second network element. A model corresponds to a network element, for example, indicating that the model is deployed within the network element or that the network element can perform inference based on the model. Alternatively, the second network element may be the coordinator (or collaborator) of model inference. The coordinator may not deploy a model and is responsible for coordinating model inference or training among various participants (including the initiator and other participants). In this case, the second network element does not deploy a corresponding model, and the one or more models include models corresponding to other participants in addition to the model corresponding to the first network element. In this case, the second model may be, for example, the model corresponding to the third network element, or the second model may replace the model described as the third network element, for example, the local model of the third network element.

[0010] In the embodiment of the present application, since each network element (such as the first network element, or the first network element and the second network element) participating in the model reasoning can be clearly used to determine the time of the reasoning data. For example, the time represents the time of collecting or generating the reasoning data, then each network element can collect or generate the reasoning data at the same or approximate time, so that the time information represented by the reasoning data of each network element is the same or approximate. Compared with the current method in which each participant performs model reasoning based on their own collected data, in the embodiment of the present application, each network element participating in the model reasoning can perform reasoning based on reasoning data with relatively consistent collection or generation time, thereby improving the accuracy of the reasoning results.

[0011] In one possible implementation, the method further includes: the second network element obtains an inference result of the second model based on the second data and the second model, wherein the second data is determined based on the time used to determine the inference data, and wherein the third information is determined based on the inference result of the second model.

[0012] In the above embodiment, if the second network element is a participant in model inference, the second network element may also determine the inference data (e.g., the second data) based on the time used to determine the inference data, thereby determining the inference result of the second model based on the second data. In this way, both participants (e.g., the first network element and the second network element) can collect or generate inference data at the same or similar time, thereby improving the accuracy of the inference result.

[0013] In one possible implementation, the method further includes: receiving third information. Optionally, the third information may be received from a third network element, for example, the third network element determines an inference result of the second model based on the second data and the second model; and the third network element sends the third information to the second network element.

[0014] In the above embodiment, if the second network element can be the coordinator of model inference and the third network element can be a participant in the model inference, the third network element can determine the inference data (e.g., the second data) based on the time used to determine the inference data, thereby determining the inference result of the second model based on the second data. In this way, both participants (e.g., the first network element and the third network element) can determine that the inference data was collected or generated at the same or similar time, thereby improving the accuracy of model inference.

[0015] In one possible embodiment, the first information indicates the time used to determine the inference data, including: the first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates the moment when the first data (or inference data) is collected or generated, and the first time window is a time window for collecting or generating the first data (or inference data).

[0016] In the above-mentioned embodiment, the first information indicates a variety of ways to determine the time for inference data, such as indicating a timestamp and / or indicating a time window. If the first information indicates a timestamp, the moment for collecting or generating inference data can be more strictly limited, thereby facilitating each network element involved in model inference to determine data that is closer in time, thereby improving the accuracy of model inference. If the first information indicates a time window, there is a certain degree of leniency in the time for collecting or generating inference data, which helps to tolerate implementation errors between different devices.

[0017] In one possible implementation, the information of the first timestamp indicates the moment when the inference data is collected or generated, including: the information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of the inference data starts, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data starts relative to the first moment; or, the first moment is the moment when the collection or generation of the inference data ends, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data ends relative to the first moment.

[0018] In the above embodiment, when indicating the first timestamp, in addition to indicating the first moment, the time offset can also be indicated, so that the time of the inference data allows a certain time offset, which is equivalent to taking into account the operational deviation of each network element, and increasing the probability that each network element participating in the model inference can determine the appropriate inference data based on the information of the first timestamp.

[0019] In a possible implementation, the start time of the first time window is the start time of collecting or generating the inference data, and the end time of the first time window is the end time of collecting or generating the inference data.

[0020] For example, the start time of the first time window is the time when inference data collection begins, and the end time of the first time window is the time when inference data collection ends. Alternatively, the start time of the first time window is the time when inference data generation begins, and the end time of the first time window is the time when inference data generation ends. Alternatively, the start time of the first time window is the time when inference data generation begins, and the end time of the first time window is the time when inference data collection ends, etc. The specific content of the first time window is not limited.

[0021] In the above implementation, the information indicating the first time window allows data collected or generated within a certain time range to be used as inference data, which is equivalent to taking into account the operational deviations of each network element, thereby increasing the probability that each network element participating in model inference can determine appropriate inference data based on the information of the first time window.

[0022] In one possible implementation, the information of the first time window indicates at least one of the following: the length of the first time window, the start time of the first time window, the end time of the first time window, or a first period, where the first period is the period of multiple time windows included in the first time window.

[0023] In the above embodiments, various implementations of the content of the first time window information are provided. When the first time window information indicates a first period, multiple recurring first time windows can be indicated by indicating the first period, without having to indicate the multiple recurring first time windows separately, thereby relatively reducing signaling overhead.

[0024] In one possible implementation, the first information also indicates at least one of the following: the moment when the first network element performs model reasoning (such as the moment when reasoning based on the first model is started); a first condition, the first condition is used to trigger the first network element to perform reasoning based on the model (or can be described as the first condition is used to trigger the first network element to perform reasoning based on the model of the first network element); or, first indication information, the first indication information is used to instruct the first network element to start model reasoning (or can be described as the first network element immediately performing model reasoning).

[0025] In the above embodiment, the first information indicates the time, first condition or first indication information for the first model to perform inference, which is beneficial for each network element for model inference to perform model inference at the same or similar time, and facilitates the second network element to obtain the inference results of the models corresponding to other network elements in a timely manner, and also facilitates the timely determination of the first inference result.

[0026] In one possible implementation, the first information indication is used to determine the time for inference data, including: the first information includes one or more of information on the moment when the first network element performs model inference, information on the first condition, or first indication information, and one or more indications are used to determine the time for inference data.

[0027] In the above implementation, on the one hand, a method for indicating the time for determining the inference data is provided, and on the other hand, the first information indicates the time for determining the inference data by indicating other information, which is conducive to reducing signaling overhead.

[0028] In one possible implementation, the method further includes: the first network element sending the following information to the second network element: first time information, the first time information indicating the time when the first data is actually collected or generated, the first data being the inference data of the first model; and / or, the sequence information of multiple inference results included in the inference result of the first model.

[0029] In the above implementation, the second network element can verify the suitability of the first data based on the first time information to ensure the rationality of the inference data used by each network element participating in model inference, thereby ensuring the accuracy of model inference. Furthermore, the second network element specifies the order of the multiple inference results corresponding to the first network element, allowing the second network element to more accurately determine the first inference result based on the order and the multiple inference results.

[0030] In one possible implementation, the method further includes: receiving information about at least one data from the first network element, the at least one data being determined based on the first information, and the at least one data including the first data; and sending fourth information to the first network element, the fourth information instructing the first network element to use the first data as inference data for the first model.

[0031] In the above embodiment, the second network element can filter out the inference data (such as the first data) ultimately used for inference from the data determined by the first network element based on the first information. In this way, the second network element can filter out the inference data that is closer in collection or generation time, which is conducive to improving the accuracy of model inference.

[0032] In one possible implementation, the first data information includes at least one of the following: information about the first data number; data feature information indicating attributes of the first data; first time information indicating the time when the first data was actually collected or generated; first location information indicating the location where the first data was collected; or access type information indicating the access type corresponding to the first data. The information about each data in the at least one data item can refer to the content of the first data information and is not listed here one by one.

[0033] The above implementations provide multiple ways to implement the information of the first data. On the one hand, the first network element does not need to send the first data itself to the second network element, ensuring data security. On the other hand, the second network element can select more appropriate first data as inference data based on this information.

[0034] In one possible implementation, the first data satisfies at least one of the following: the difference between the number of the first data and the number of the inference data corresponding to one or more models is less than or equal to a first threshold; the difference between the actual collection or generation time of the first data and the time of the inference data corresponding to one or more models is less than or equal to a second threshold; the distance between the location where the first data is collected and the location of the inference data corresponding to one or more models is less than or equal to a third threshold; or, the access type corresponding to the first data is the same as the access type corresponding to the inference data corresponding to one or more models.

[0035] In the above embodiment, the second network element may determine the first data from the at least one data based on at least one item of the data's corresponding number, time, location, or access type. This facilitates the second network element to determine inference data with similar or identical numbers, time, location, or access types for each network element participating in model inference, thereby improving the accuracy of the inference results.

[0036] In a second aspect, an embodiment of the present application provides a communication method. The method can be applied to a second network element. The implementation method of the second network element can refer to the content of the implementation method of the network element discussed in the first aspect above, and will not be listed here. The method includes: sending first information, the first information indicating the time for determining inference data, the inference data is used for inference based on a first model, and the first model is associated with one or more models; receiving second information, the second information indicating the inference result based on the first model; determining the first inference result based on the second information and third information, the third information indicating the inference result of the second model, and the second model is one of the one or more models.

[0037] In one possible implementation, the method further includes: obtaining an inference result of the second model based on second data and a second model, wherein the second data is determined based on a time used to determine the inference data; and obtaining third information according to the inference result of the second model.

[0038] In one possible implementation, the first information indicates the time used to determine the inference data, including: the first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates the moment when the inference data is collected or generated, and the first time window is a time window for collecting or generating the inference data.

[0039] In one possible implementation, the information of the first timestamp indicates the moment when the inference data is collected or generated, including: the information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of the inference data starts, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data starts relative to the first moment; or, the first moment is the moment when the collection or generation of the inference data ends, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data ends relative to the first moment.

[0040] In a possible implementation, the start time of the first time window is the start time of collecting or generating the inference data, and the end time of the first time window is the end time of collecting or generating the inference data.

[0041] In one possible implementation, the information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; the end time of the first time window; or the first period, which is the period of multiple time windows included in the first time window.

[0042] In one possible implementation, the first information also indicates at least one of the following: the moment when the first network element performs model inference; a first condition, the first condition is used to trigger the first network element to perform inference based on the model; or, first indication information, the first indication information is used to instruct the first network element to start model inference.

[0043] In one possible implementation, the first information indication is used to determine the time of inference data, including: information of the moment when the first network element performs model inference, information of the first condition, or one or more of the first indication information, and one or more indications are used to determine the time of inference data.

[0044] In one possible implementation, the method further includes: receiving the following information: first time information, the first time information indicating the time when the first data is actually collected or generated, the first data being the inference data of the first model; and / or, sequence information of multiple inference results included in the inference result of the first model.

[0045] In one possible implementation, the method further includes: receiving information about at least one data from the first network element, the at least one data being determined based on the first information, and the at least one data including the first data; and sending fourth information to the first network element, the fourth information instructing the first network element to use the first data as inference data for the first model.

[0046] In one possible embodiment, the information of the first data includes at least one of the following: information on the number of the first data; data feature information, the data feature information indicates the attributes of the first data; first time information, the first time information indicates the time when the first data is actually collected or generated; first location information, the first location information indicates the location where the first data is collected; or access type information, the access type information indicates the access type corresponding to the first data.

[0047] In one possible implementation, the first data satisfies at least one of the following: the difference between the number of the first data and the number of the inference data used by one or more models is less than or equal to a first threshold; the difference between the actual collection or generation time of the first data and the time of the inference data corresponding to one or more models is less than or equal to a second threshold; the distance between the location where the first data is collected and the location of the inference data corresponding to one or more models is less than or equal to a third threshold; or, the access type corresponding to the first data is the same as the access type corresponding to the inference data corresponding to one or more models.

[0048] In a third aspect, an embodiment of the present application provides a communication method. The method can be applied to a first network element. The implementation method of the first network element can refer to the implementation method of the network element discussed in the first aspect above, and will not be listed here. The method includes: receiving first information, the first information indicating the time for determining inference data, and the inference data is used for inference based on a first model; obtaining an inference result of the first model based on the first data and the first model, the first data being the inference data determined based on the first information; and sending second information to a second network element, the second information indicating the inference result of the first model. The first model is associated with one or more models.

[0049] In one possible implementation, the first information indicates the time used to determine the inference data, including: the first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates the moment when the inference data is collected or generated, and the first time window is a time window for collecting or generating the inference data.

[0050] In one possible implementation, the information of the first timestamp indicates the moment when the inference data is collected or generated, including: the information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of the inference data starts, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data starts relative to the first moment; or, the first moment is the moment when the collection or generation of the inference data ends, and the time offset indicates the maximum deviation of the moment when the collection or generation of the inference data ends relative to the first moment.

[0051] In a possible implementation, the start time of the first time window is the start time of collecting or generating the inference data, and the end time of the first time window is the end time of collecting or generating the inference data.

[0052] In one possible implementation, the information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; the end time of the first time window; or the first period, which is the period of the first time window.

[0053] In one possible implementation, the first information also includes at least one of the following: the moment when the first network element performs model inference; a first condition, the first condition is used to trigger the first network element to perform inference based on the model; or, first indication information, the first indication information is used to instruct the first network element to start model inference.

[0054] In one possible implementation, the first information indication is used to determine the time for inference data, including: the first information includes one or more of information on the moment when the first network element performs model inference, information on the first condition, or first indication information, and one or more indications are used to determine the time for inference data.

[0055] In one possible implementation, the method further includes: sending the following information to the second network element: first time information, the first time information indicating the time when the first data is actually collected or generated, the first data being the inference data of the first model; and / or, sequence information of multiple inference results included in the inference result of the first model.

[0056] In one possible implementation, the method further includes: sending information of at least one data to the second network element, the at least one data is determined based on the first information, and the at least one data includes the first data; and receiving fourth information, the fourth information indicating the use of the first data for inference.

[0057] In one possible implementation, the information of the first data includes at least one of the following: information on the number of the first data; data feature information, the data feature information indicates the attributes of the first data; first time information, the first time information indicates the time when the first data is actually collected or generated; first location information, the first location information indicates the location where the first data is collected; or access type information, the access type information indicates the access type corresponding to the first data.

[0058] In a fourth aspect, embodiments of the present application provide a communication method. This method can be applied to a second network element. The implementation of the second network element can refer to the implementation of the network element discussed in the first aspect above and is not further detailed here. The method includes: sending first information indicating a time for determining training data, the training data being used to train a first model, the first model being associated with one or more models.

[0059] In the embodiment of the present application, since each network element participating in the model training (for example, the first network element, or the first network element and the second network element) can clearly determine the time of the training data, for example, the time represents the time of collecting or generating the training data, then each network element can determine the training data with a relatively consistent time of collecting the data or generating the data, and can also perform model training based on the training data with a relatively consistent time of collecting or generating the data. Compared with the current practice of each participant performing model training based on their own collected data, the network elements participating in the model training in the embodiment of the present application can perform model training based on training data with a relatively consistent time of collecting or generating the data, thereby improving the accuracy of model training.

[0060] In one possible implementation, the method further includes: receiving second information sent from a first network element, the second information indicating a training result of a first model of the first network element; determining a first training result based on the training result of the first model and third information, the third information indicating a training result of a second model, the second model belonging to one or more models.

[0061] In a possible implementation, the method further includes: determining a training result of the second model based on the second data and the second model. Alternatively, receiving third information.

[0062] In one possible embodiment, the first information indicates information used to determine a first timestamp and / or information of a first time window, wherein the information of the first timestamp and / or the information of the first time window indicates a time for determining training data, wherein the first timestamp indicates a moment when the first data is collected or generated, and the first time window is a time window for collecting or generating the first data.

[0063] In one possible implementation, the information of the first timestamp indicates the moment when the first data is collected or generated, including: the information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of the first data starts, and the time offset is the maximum deviation of the moment when the collection or generation of the first data starts relative to the first moment; or, the first moment is the moment when the collection or generation of the first data ends, and the time offset is the maximum deviation of the moment when the collection or generation of the first data ends relative to the first moment.

[0064] In a possible implementation, the start time of the first time window is the time when the collection or generation of the first data starts, and the end time of the first time window is the time when the collection or generation of the first data ends.

[0065] In a possible implementation, the information of the first time window includes at least one of the following: the length of the first time window; the start time of the first time window; the end time of the first time window; or the first period, which is the period of the first time window.

[0066] In one possible implementation, the first information also includes at least one of the following: the moment when the first network element performs model training; a first condition, the first condition is used to trigger the first network element to perform training based on the model; or, first indication information, the first indication information is used to instruct the first network element to start model training.

[0067] In one possible implementation, the first information indication is used to determine the time of training data, including: the first information includes information of the moment when the first network element performs model training, information of the first condition, or one or more of the first indication information, one or more of which are used to determine the time of training data.

[0068] In one possible implementation, the method further includes: sending one or more of the following information to the first network element: information on the moment when the first data is actually collected or generated; the time window when the first data is actually collected or generated; or sequence information of multiple training results included in the first training result.

[0069] In one possible implementation, the method further includes: sending information of at least one data to the second network element, the at least one data being determined based on a time used to determine training data, and the at least one data including first data; and receiving fourth information, the fourth information indicating the use of the first data for training.

[0070] In one possible implementation, the information of the first data includes at least one of the following: the number of the first data; data characteristic information, the data characteristic information indicates the characteristics of the first data; first time information, the first time information indicates the time when the first data is actually collected or generated; first location information, the first location information indicates the location where the first data is collected; or access type information, the access type information indicates the access type corresponding to the first data.

[0071] In one possible implementation, the first data satisfies at least one of the following conditions: the difference between the number of the first data and the number of the training data corresponding to one or more models is less than or equal to a first threshold; the difference between the actual collection or generation time of the first data and the time of the training data corresponding to one or more models is less than or equal to a second threshold; the distance between the location where the first data is collected and the location of the training data corresponding to one or more models is less than or equal to a third threshold; or, the access type corresponding to the first data is the same as the access type corresponding to the training data corresponding to one or more models.

[0072] In a possible implementation, the method further includes: receiving fifth information, where the fifth information indicates a first training result; and determining, based on the first training result, whether to end the training of the multiple models or to continue the training of the multiple models.

[0073] The beneficial effects in any possible implementation of the fourth aspect can refer to the beneficial effects discussed in any possible implementation of the first aspect, and will not be listed one by one here.

[0074] In a fifth aspect, an embodiment of the present application provides a communication method. The method can be applied to a first network element. The implementation method of the first network element can refer to the implementation method of the network element discussed in the first aspect above, and will not be listed here. The method includes: receiving first information, the first information indicating the time for determining training data, and the training data is used to train a first model; obtaining a training result of the first model based on the first data and the first model, the first data is training data determined according to the first information, and the first model belongs to multiple models; sending second information to a second network element, the second information indicating the training result of the first model. Wherein, the first model is associated with one or more models.

[0075] In one possible implementation, the first information indicates the time for determining the training data, including: the first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates the moment when the training data is collected or generated, and the first time window is a time window for collecting or generating the training data.

[0076] In one possible implementation, the information of the first timestamp indicates the moment when the training data is collected or generated, including: the information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of the training data starts, and the time offset indicates the maximum deviation of the moment when the collection or generation of the training data starts relative to the first moment; or, the first moment is the moment when the collection or generation of the training data ends, and the time offset indicates the maximum deviation of the moment when the collection or generation of the training data ends relative to the first moment.

[0077] In a possible implementation, the start time of the first time window is the start time of collecting or generating the training data, and the end time of the first time window is the end time of collecting or generating the training data.

[0078] In one possible implementation, the information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; the end time of the first time window; or the first period, which is the period of the first time window.

[0079] In one possible implementation, the first information also includes at least one of the following: the moment when the first network element performs model training; a first condition, the first condition is used to trigger the first network element to perform training based on the model; or, first indication information, the first indication information is used to instruct the first network element to start model training.

[0080] In one possible implementation, the first information indication is used to determine the time of training data, including: information of the moment when the first network element performs model training, information of the first condition, or one or more of the first indication information, and one or more indications are used to determine the time of training data.

[0081] In a possible implementation, the method further includes: sending the following information to the second network element: first time information, the first time information indicating the time when the first data is actually collected or generated, the first data being the training data of the first model; and / or, sequence information of multiple training results included in the training results of the first model.

[0082] In a possible implementation, the method further includes: sending information of at least one data to the second network element, the at least one data is determined based on the first information, and the at least one data includes the first data; and receiving fourth information, the fourth information indicating the use of the first data for training.

[0083] In one possible implementation, the information of the first data includes at least one of the following: information on the number of the first data; data feature information, the data feature information indicates the attributes of the first data; first time information, the first time information indicates the time when the first data is actually collected or generated; first location information, the first location information indicates the location where the first data is collected; or access type information, the access type information indicates the access type corresponding to the first data.

[0084] In a sixth aspect, embodiments of the present application provide a communication system. The communication system can implement the method described in the first aspect and any possible implementation. Optionally, the communication system can include the first network element and the second network element involved in the first aspect.

[0085] In a seventh aspect, an embodiment of the present application provides a communication device. The communication device may be the second network element in the second aspect above, or a chip system (or chip) or other functional module configured in the second network element. The contents of other functional modules can refer to the contents of other functional modules discussed in the first aspect above, and are not listed here. The communication device includes corresponding means (means) or modules for executing the second aspect above or any possible implementation method. For example, the communication device includes a processing module (sometimes also referred to as a processing unit), and a transceiver module (sometimes also referred to as a transceiver unit).

[0086] For example, the transceiver module is configured to send first information and receive second information, and the processing module is configured to determine the first inference result based on the second information.

[0087] In a possible implementation, the communication device may also execute any possible implementation in the third aspect above, which will not be listed one by one here.

[0088] In an eighth aspect, an embodiment of the present application provides a communication device. The communication device may be the first network element in the third aspect above, or a chip system (or chip) or other functional module configured in the first network element. The contents of other functional modules can refer to the contents of other functional modules discussed in the first aspect above, and are not listed here. The communication device includes corresponding means (means) or modules for executing the third aspect above or any possible implementation method. For example, the communication device includes a processing module (sometimes also referred to as a processing unit), and a transceiver module (sometimes also referred to as a transceiver unit).

[0089] For example, the transceiver module is used to receive the first information and send the second information. The processing module is used to obtain the inference result of the first model based on the first data and the first model.

[0090] In a possible implementation, the communication device may also execute any possible implementation in the third aspect above, which will not be listed one by one here.

[0091] In a ninth aspect, an embodiment of the present application provides a communication device. The communication device may be the second network element in the fourth aspect above, or a chip system (or chip) or other functional module configured in the second network element. The contents of other functional modules can refer to the contents of other functional modules discussed in the first aspect above, and are not listed here. The communication device includes corresponding means (means) or modules for executing the fourth aspect or any possible implementation method. For example, the communication device includes a processing module (sometimes also referred to as a processing unit), and a transceiver module (sometimes also referred to as a transceiver unit).

[0092] For example, the transceiver module is used to send the first information under the control of the processing module.

[0093] In a possible implementation, the communication device can also execute any possible implementation in the fourth aspect above, which will not be listed one by one here.

[0094] In the tenth aspect, an embodiment of the present application provides a communication device. The communication device can be the first network element in the fifth aspect above, or a chip system (or chip) or other functional module configured in the first network element. The contents of other functional modules can refer to the contents of other functional modules discussed in the first aspect above, and are not listed here. The communication device includes corresponding means (means) or modules for executing the fifth aspect or any possible implementation method. For example, the communication device includes a processing module (sometimes also referred to as a processing unit), and a transceiver module (sometimes also referred to as a transceiver unit).

[0095] For example, the transceiver module is used to receive the first information and send the second information. The processing module is used to determine the inference result of the first model.

[0096] In a possible implementation, the communication device can also execute any possible implementation in the fifth aspect above, which will not be listed one by one here.

[0097] In an eleventh aspect, an embodiment of the present application provides a communication system. The communication system includes any communication device described in the seventh aspect and any communication device described in the eighth aspect. Alternatively, the communication system includes any communication device described in the ninth aspect and any communication device described in the tenth aspect.

[0098] In a twelfth aspect, an embodiment of the present application provides a communication device, comprising: a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to execute the program instructions in the memory to implement any one of the methods according to the second to fifth aspects.

[0099] Optionally, the wireless communication device further includes a communication interface, and the processor may be coupled to the communication interface, or the processor may be independently provided with the communication interface.

[0100] In one implementation, when the communication apparatus is a wireless communication device, the communication interface may be a transceiver or an input / output interface. Alternatively, the transceiver may be a transceiver circuit. Alternatively, the input / output interface may be an input / output circuit.

[0101] In another implementation, when the communication device is a chip or a chip system, the communication interface may be an input / output interface, an interface circuit, an output circuit, an input circuit, a pin, or related circuits on the chip or chip system. The processor may also be embodied as a processing circuit or a logic circuit.

[0102] Optionally, the communication device further includes other components, such as an antenna, an input / output module or an interface, etc. These components may be hardware, software, or a combination of software and hardware.

[0103] In a thirteenth aspect, an embodiment of the present application provides a communication device. The communication device includes: a processing circuit and an interface circuit; wherein: the interface circuit is configured to couple with a memory external to the communication device and provide a communication interface for the processing circuit to access the memory; and the processing circuit is configured to execute program instructions in the memory to implement the method according to any one of aspects 2 to 5.

[0104] In a specific implementation, the communication device may be a chip, the input circuit may be an input pin, the output circuit may be an output pin, and the processing circuit may be a transistor, a gate circuit, a trigger, or various logic circuits. The input signal received by the input circuit may be, for example, but not limited to, received and input by a receiver, and the signal output by the output circuit may be, for example, but not limited to, output to and transmitted by a transmitter. The input circuit and the output circuit may be the same circuit, which functions as an input circuit and an output circuit at different times. The embodiments of the present application do not limit the specific implementation of the processor and various circuits.

[0105] In one implementation, the communication device may be a wireless communication device, that is, a computer device that supports wireless communication functions. Specifically, the wireless communication device may be a terminal such as a smartphone, or a wireless access network device such as a base station. The system chip may also be referred to as a system on chip (SoC), or simply as an SoC chip. The communication chip may include a baseband processing chip and a radio frequency processing chip. The baseband processing chip is sometimes also referred to as a modem or baseband chip. The radio frequency processing chip is sometimes also referred to as a radio frequency transceiver or radio frequency chip. In a physical implementation, some or all of the chips in the communication chip may be integrated inside the SoC chip. For example, the baseband processing chip is integrated into the SoC chip, and the radio frequency processing chip is not integrated with the SoC chip. The interface circuit may be the radio frequency processing chip in the wireless communication device, and the processing circuit may be the baseband processing chip in the wireless communication device.

[0106] In another implementation, the communication device may be a component of a wireless communication device, such as an integrated circuit product such as a system-on-chip (SoC) or a communication chip. The interface circuit may be an input / output interface, interface circuit, output circuit, input circuit, pin, or related circuit on the chip or chip system. The processor may also be embodied as a processing circuit or a logic circuit.

[0107] In a fourteenth aspect, an embodiment of the present application provides a chip system. The chip system includes a processor and an interface. The processor is configured to call and execute instructions from the interface. When the processor executes the instructions, the method described in any one of aspects 1 to 5 above is implemented.

[0108] In a fifteenth aspect, an embodiment of the present application provides a computer-readable storage medium for storing a computer program or instruction, which, when executed, implements the method described in any one of the first to fifth aspects above.

[0109] In a sixteenth aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when executed on a computer, implements the method described in any one of the first to fifth aspects above.

[0110] Regarding the beneficial effects of the second to third aspects, and the fifth to sixteenth aspects, reference may be made to the beneficial effects discussed in the first or fourth aspect, and they will not be listed here. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] Figure 1 is a structural diagram of the model;

[0112] FIG2 is a schematic diagram of data distribution for vertical federated learning applicable to an embodiment of the present application;

[0113] FIG3 is a schematic diagram of a scenario applicable to an embodiment of the present application;

[0114] FIG4 is a schematic diagram of another scenario provided in an embodiment of the present application;

[0115] FIG5 is a schematic diagram of an intelligent network architecture based on NWDAF provided in an embodiment of the present application;

[0116] FIG6 is a schematic diagram of another scenario provided in an embodiment of the present application;

[0117] Figures 7 to 10 are schematic diagrams of several communication methods provided in embodiments of the present application;

[0118] 11 to 13 are schematic structural diagrams of several communication devices provided in embodiments of the present application. DETAILED DESCRIPTION

[0119] The specific implementation of the present application is described below with reference to the drawings in the embodiments of the present application.

[0120] Below, some terms used in the embodiments of the present application are explained to facilitate understanding by those skilled in the art.

[0121] 1. Model

[0122] It is a specific implementation of one or more functions, characterizing the mapping relationship between the input and output of the model. In the fields of machine learning (ML) and artificial intelligence (AI), a model can be understood as an algorithm or system that can make predictions or perform tasks after training and learning based on input data. Models include, for example, ML models, AI models, algorithms, features or functions. The AI ​​model can be at least one of a linear regression model, a logistic regression model, a decision tree model, a support vector machine (SVM), a neural network model, a clustering model, a Bayesian network, a Q learning model, a generative adversarial network, or other machine learning models, without limitation. The neural network model can be, for example, a multilayer perceptron (MLP), a deformation model (or a conversion network or conversion model, etc.) (transformer network), a convolutional neural network (CNN), or an attention mechanism, without specific limitation.

[0123] Taking the neural network model as an example, a model may include at least one layer, and the "layer" may include a "network layer". Each "network layer" may contain at least one node, which may also be called a "neuron". Please refer to Figure 1, which is a schematic diagram of the structure of the model. Taking the model shown in Figure 1 as an example, there is an input layer, a hidden layer, and an output layer. Optionally, the model may also include a loss layer. Any layer involved here can be regarded as a network layer. At least one operator may be included between network layers, such as a convolution operator, a fully connected operator, etc.

[0124] Neurons in a network layer are connected to neurons in adjacent layers through operators, and each connection can be considered an operation. Taking the connection between the input layer and hidden layer 1 shown in Figure 1 as an example, the input layer and hidden layer 1 are fully connected operators, meaning that every neuron in the input layer is connected to every neuron in hidden layer 1. This fully connected operator or operation can also be described as a "fully connected layer." In this context, the terms "layer" and "operator" in a fully connected layer are equivalent and interchangeable, and can be referred to as an "operator layer." In some cases, an "operator layer" can also include the previous and / or next adjacent network layers. Still using the connection between the input layer and hidden layer 1 shown in Figure 1 as an example, this "fully connected layer" can also include the "input layer" and / or "hidden layer 1." In addition to fully connected layers, a model can also include convolutional layers. For example, in a convolutional neural network, a convolution operator can include one or more convolution kernels. Optionally, these one or more convolution kernels can be divided into at least one convolution kernel group, and a convolution kernel group can include at least one convolution kernel. A "convolution kernel group" may also be referred to as a "filter." A model may include one or more operators. A substructure (or submodule) of a model may include one or more operators.

[0125] Continuing to refer to Figure 1, the AI / ML model includes five network layers, namely the input layer, hidden layer 1, hidden layer 2, output layer, and loss layer. Among them, the loss layer corresponds to the cross entropy loss function, for example. The circles in Figure 1 represent neurons, and the lines between the circles between the network layers represent connections. It can be seen that the input layer and hidden layer 1 are fully connected operators.

[0126] Optionally, the fully connected operator can also be represented by a “fully connected layer”.

[0127] The models involved in the various embodiments of the present application can also be called distributed models, which is not specifically limited.

[0128] 2. Data, including training data and inference data

[0129] Training data is used for model training. Model training can also be described as model learning. Model inference can also be referred to as model usage. Alternatively, training data can serve as input to the model during training, or training data can be used to determine the input to the model during training. Similarly, inference data is used for model inference. Alternatively, inference data can serve as input to the model during training, or inference data can be used to determine the input to the model during inference.

[0130] Training data includes structured data and / or unstructured data, without limitation. Structured data can include at least one item such as a table or database. Unstructured data can include at least one item such as text, images, audio, or video. The format of inference data can also refer to the format of training data and will not be listed here.

[0131] Both training data and inference data can be divided into two dimensions (or spaces), one dimension is the object space and the other dimension is the feature space. The object space represents the set of objects corresponding to the data, and the feature space represents the set of features used to participate in model inference or training. For example, in the training phase of the model, the object space can indicate which objects (such as users) are trained or learned. In this case, the object space can also be called the sample space. The feature space can indicate which features of the objects are trained or learned. Alternatively, in the inference phase of the model, the object space can indicate which objects (such as users) are inferred, or based on which objects the inference is performed. The feature space can indicate which features of the objects are inferred, or based on which features of the objects the inference is performed.

[0132] Features describe data attributes and can be used to describe key aspects of the data, enabling models to better learn or reason. Features can be obtained by preprocessing raw data. Preprocessing can include, but is not limited to, sorting, vector representation, matrix representation, statistical averaging, feature extraction, or dimensionality reduction. During the model training phase, labels are associated with features; labels represent the desired predictions of the model.

[0133] For example, consider training a model for image classification to distinguish between images of cats and dogs. The training data for this model can be a set of images labeled as either cats or dogs. Each image in this set consists of two components: features and labels. In this example, features are attributes that describe the image, such as pixel values, color histograms, or texture features. The label is the result of classifying the image, i.e., whether the image contains a cat or a dog.

[0134] For example, please refer to Table 1 below, which is an example of inference data.

[0135] Table 1

[0136] As shown in Table 1 above, the object space corresponding to the inference data includes user 1 and user 2, and the feature space includes wireless access type, reference signal received power, reference signal received quality, and wireless access network throughput. Table 1 is an example of inference data and does not limit the specific form and content of inference data. Optionally, the content shown in Table 1 can also be used as an example of training data.

[0137] 3. Distributed Learning

[0138] This method, also known as distributed training, refers to a learning method that combines data from multiple participants to perform model learning. Specifically, multiple participants deploy their own models. When training these models, each participant trains their own models based on their own data, and then conducts joint training based on the intermediate results of the training. In this way, the participants can achieve joint training without sharing data. A typical example of distributed learning is federated learning (FL). Federated learning is a machine learning framework that can effectively help multiple users use data and conduct machine learning modeling while meeting the requirements of user privacy protection, data security, and government regulations. As a distributed machine learning paradigm, federated learning can effectively solve the problem of data silos, perform joint modeling without sharing user data, and thus technically break down data silos and achieve artificial intelligence (AI) collaboration. Distributed learning can be applied when two training entities cannot share their models. For example, the input data used to train the model cannot be shared (to avoid bulk data collection or privacy issues) due to privacy between the application function (AF) and the network data analytics function (NWDAF), or due to privacy between multi-vendor NWDAFs.

[0139] Based on the characteristics of the data provided by each participant, federated learning can be divided into three categories: horizontal federated learning, transfer learning, and vertical federated learning (VFL).

[0140] In horizontal federated learning, the object overlap corresponding to the data is low (or can be described as less overlap), but the feature overlap is high (or can be described as more overlap). For example, the participants in horizontal federated learning include Participant 1 and Participant 2. The data provided by Participant 1 includes Feature 1 and Feature 2 of User 1, and the data provided by Participant 2 includes Feature 1 and Feature 2 of User 2. Transfer learning is a method of applying the model of a source task in a source domain to a target task in a target domain. For example, Task A (source task) is to classify cats and tigers in pictures, and Task B (target task) is to distinguish the lengths of cats and tigers in pictures. Traditional machine learning methods can only train two different models for these two completely different tasks. However, using transfer learning, we can fully utilize the model of Task A and fine-tune it based on the model of Task A to obtain a model that is applicable to Task B.

[0141] As a machine learning technique, vertical federated learning can be used to address model training and inference challenges when participants are unwilling to share their original datasets. It is suitable for scenarios where participants (or participants) have high overlap in their objects (such as samples) but low overlap in their features. Vertical federated learning combines different datasets (or features) of shared samples from multiple participants for federated learning. This means that the data (or features) of each participant are partitioned vertically, hence the name.

[0142] For example, please refer to Figure 2, which shows a data distribution diagram for a vertical federated learning scenario. Figure 2 takes the federated learning scenario as an example, with participants A and B as the participants. As shown in Figure 2, the sample spaces corresponding to participants A and B have a high degree of overlap, but the feature spaces have a low degree of overlap. For example, participant A's dataset contains data for user 1, user 2, user 3, and user 4, where each user's data includes features 1, 2, 3, 4, and 5. Participant B's dataset contains data for users 1, 2, 3, 4, and 5, where each user's data includes features 6, 7, 8, 9, and 10. This shows that the feature overlap between participants A and B is small, but the samples are mostly identical. When conducting vertical federated learning, participant A and participant B can select the features (i.e., feature 1, feature 2, feature 3, feature 4, feature 5, feature 6, feature 7, feature 8, feature 9, and feature 10) corresponding to the same samples (i.e., user 1, user 2, user 3, and user 4) from their respective data sets, as well as the labels corresponding to these features, to conduct model training for vertical federated learning.

[0143] All parties involved in distributed learning can be regarded as participants (or participants), and the participants can be divided into master participants and slave participants. There can be one or more slave participants, and there is no limitation on this. The master participant has the labels of the data required for the distributed learning task, and optionally has some of the data features required for the distributed learning task. The slave participant has some or all of the data features required for the vertical federated learning task. The master participant can also be called an active participant, an active participant (active participant) or an active client (active client), etc. The slave participant can also be called a passive participant (passive participant) or a passive client (passive client), etc.

[0144] Distributed learning also includes a coordinator. This can also be called a collaborator, coordinator, or server. A coordinator is a trusted party responsible for maintaining the distributed learning process, authorizing the entry and removal of distributed learning members, and optionally distributing encryption keys and decrypting intermediate information. This role is typically performed by a third party independent of the task or by a credible organization within the industry.

[0145] In addition, the participant who initiates the distributed learning may be called an initiator. The initiator may be a master participant, a slave participant, or a coordinator in the distributed learning, without limitation.

[0146] The main feature of distributed learning (such as VFL joint model training) is the presence of multiple models (such as ML models), each of which is associated with a different training entity (such as a network element). Different training entities include (or correspond to) a master participant and at least one slave participant. The master participant and the slave participant each have their own models. The master participant's model and the slave participant's model have the same model objective, but the input data or data type corresponding to these multiple models may be different. The goal of multiple models is to predict the same output, such as the service experience.

[0147] Optionally, distributed learning corresponds to a distributed learning identification (ID). This distributed learning identification can also be called a global model ID. This distributed learning ID is associated with the distributed learning process and can be considered a unique identifier for the distributed learning process. The distributed learning ID can also be used to uniquely identify the model of the master participant and the model of the slave participant.

[0148] Corresponding to distributed learning is distributed reasoning (or distributed inference). Distributed reasoning refers to the joint reasoning of various models trained in distributed manner. In other words, distributed learning corresponds to the training phase of the model, and distributed reasoning corresponds to the reasoning phase of the model. The parties involved in distributed reasoning can be some or all of the parties involved in distributed learning. All parties involved in distributed reasoning can also be considered participants. Participants can also be divided into master participants, slave participants, and coordinators. The participant used to initiate the distributed reasoning process can be considered the initiator. The initiator of distributed reasoning and the initiator of distributed learning may be the same participant or different participants, and there is no specific limitation on this.

[0149] The characteristics of distributed reasoning can refer to the characteristics of distributed learning discussed above and will not be listed here. For example, distributed reasoning also involves a master participant, a coordinator, and at least one slave participant. Optionally, distributed reasoning also corresponds to a distributed reasoning ID. This distributed reasoning identifier can also be called a global model ID. This distributed reasoning ID is associated with the distributed reasoning process and can be regarded as a unique identifier of the distributed reasoning process. The distributed reasoning ID can also be used to identify the unique identifier of the master participant's model and the slave participant's model.

[0150] The solutions involved in the various embodiments of this application can be applied to the fourth generation (4 th generation, 4G), fifth generation (5 th generation, 5G) (such as new radio (NR)), the sixth generation communication system (6 th generation, 6G) or future evolution communication systems. The technical solution provided in this application can also be applied to communication systems such as side link (SL) or non-terrestrial network (NTN), etc., without limitation. SL can also be called side communication link, side link, side link, direct link, side link or auxiliary link, etc. SL includes vehicle-to-everything (V2X) communication, etc. V2X communication may include: vehicle-to-vehicle (V2V) communication, vehicle-to-roadside infrastructure (V2I) communication, vehicle-to-pedestrian (V2P) communication, vehicle-to-network (V2N) communication, etc., without specific limitation.

[0151] Please refer to Figure 3, which is a schematic diagram of a scenario applicable to embodiments of the present application. As shown in Figure 3, the scenario includes a master participant, a slave participant, and a coordinator. Any two of the master participant, the slave participant, and the coordinator can communicate via a communication network, such as a fifth-generation communication network, a sixth-generation communication network, or a future evolution communication network.

[0152] Any participant in distributed learning or distributed reasoning (such as a master participant, a slave participant, or a coordinator) can be implemented through a network element. The network element involved in the embodiments of the present application can be implemented through a device, a client, software in a device (such as an application), a chip system in a device (such as a chip), or other functional modules (such as a program, a hardware component, or a hardware component running a program, etc.), or a device cluster, etc., without specific limitation. The client can be, for example, a program running in a terminal device, or a terminal device running a program, etc. The client can also be, for example, a program or function in a network element, etc., without limitation. For example, the network element is specifically a terminal device, a network device, a third-party server, etc., without limitation.

[0153] Both the master and slave participants have models that participate in distributed learning or reasoning. Figure 3 illustrates this using an example where the master is client 1, the slave is client 2, and the coordinator is the server. The implementation of the master, slave, and coordinator is not limited. Either the master or the slave can serve as the initiator of distributed learning or distributed reasoning, and the coordinator can coordinate the master and slave participants in distributed learning or distributed reasoning.

[0154] The terminal device mentioned above is a device with wireless transceiver functions, which can be a fixed device, a mobile device, a handheld device, a wearable device, an in-vehicle device, or a wireless device built into the above device (for example, a communication module or a chip system, etc.). The terminal device is used to connect people, objects, machines, etc., and can be widely used in various scenarios, such as but not limited to the following scenarios: cellular communication, device-to-device communication (D2D), vehicle to everything (V2X), machine-to-machine / machine-type communication (M2M / MTC), Internet of Things (IoT), virtual reality (VR), augmented reality (AR), industrial control, self-driving, remote medical, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, drones, robots and other scenarios. A terminal device may sometimes be referred to as user equipment (UE), terminal, access station, UE station, remote station, wireless communication device, or user equipment.

[0155] The network equipment includes, for example, access network equipment (or, referred to as access network devices / access network network elements), and / or core network equipment (or, referred to as core network devices / core network network elements). The access network equipment is a device with wireless transceiver functions, which is used to communicate with the terminal equipment. Access network equipment includes but is not limited to base stations (BTS, Node B, eNodeB / eNB, or gNodeB / gNB) in the communication system, transmission reception points (TRP), base stations of subsequent evolution of 3GPP, access nodes in wireless fidelity (WiFi) systems, wireless relay nodes, wireless backhaul nodes, satellites or drones, etc. The base station can be: a macro base station, a micro base station, a pico base station, a small station, a relay station, etc. Multiple base stations can support the network of the same access technology mentioned above, or they can support the network of different access technologies mentioned above. The base station can include one or more co-sited or non-co-sited transmission and receiving points. The access network device can also be a wireless controller, a centralized unit (CU), also called an aggregation unit, and / or a distributed unit (DU) in a cloud radio access network (C(R)AN) scenario. The access network device can also be a server, a wearable device, or a vehicle-mounted device. For example, the access network device in the vehicle to everything (V2X) technology can be a road side unit (RSU). The following describes the access network device by taking a base station as an example. The multiple access network devices in the communication system can be base stations of the same type or different types. The base station can communicate with the terminal device or communicate with the terminal device through a relay station. The terminal device can communicate with multiple base stations in different access technologies.

[0156] In the case where the access network device includes a CU and / or a DU. CU and DU can be understood as a division of the access network device from a logical functional perspective. CU and DU can be physically separated or deployed together, and this embodiment of the present application does not specifically limit this. A CU can be connected to a DU, or multiple DUs can share a CU. The division of CU and DU can be based on the protocol stack. One possible way is to deploy the radio resource control (RRC), service data adaptation protocol stack (SDAP) and packet data convergence protocol (PDCP) layers in the CU, and the remaining radio link control (RLC) layers, media access control (MAC) layers and physical layers in the DU. The embodiment of the present application does not completely limit the division of CU and DU according to the above-mentioned protocol stack method, and there may be other division methods, such as division according to service type.

[0157] Access network equipment may also refer to a centralized unit control plane (CU-CP) node or a centralized unit user plane (CU-UP) node, or include both the CU-CP and the CU-UP. The CU-CP is responsible for control plane functions, primarily including RRC and PDCP-C. PDCP-C is primarily responsible for encryption and decryption, integrity protection, and data transmission of control plane data. The CU-UP is responsible for user plane functions, primarily including SDAP and PDCP-U. SDAP is primarily responsible for processing core network data and mapping flows to bearers. PDCP-U is primarily responsible for encryption and decryption, integrity protection, header compression, sequence number maintenance, and data transmission of the data plane.

[0158] In different systems, CU (including CU-CP or CU-UP) or DU may have different names, but those skilled in the art will understand their meanings. For example, in an open radio access network (O-RAN) system, CU may also be referred to as O-CU (Open CU), DU may also be referred to as O-DU, CU-CP may also be referred to as O-CU-CP, and CU-UP may also be referred to as O-CU-UP.

[0159] The core network equipment is used to implement at least one of the following functions: mobility management, data processing, session management, policy and billing. The names of the devices that implement the core network functions in systems with different access technologies may be different, and this embodiment of the present application is not limited to this. Taking the 5G system as an example, the core network equipment includes: access and mobility management function (AMF), session management function (SMF), or user plane function (UPF).

[0160] At present, since the time when each participant participates in the model reasoning of the data is not taken into consideration, each participant may use data with a large time difference in collection or generation to participate in the model reasoning, which may lead to low accuracy of model reasoning.

[0161] In view of this, an embodiment of the present application provides a communication method. In this method, a coordinator or initiator (such as a second network element) can indicate the time used to determine the inference data to a participant in model inference (such as a first network element), so that all participants in model inference can use inference data generated or collected at a similar or identical time to perform model inference, which is conducive to improving the accuracy of model inference.

[0162] The communication method provided in the embodiment of the present application can be applied to the scenario involved in Figure 3 above. In addition, it can also be applied to other similar scenarios, which are described below with examples in conjunction with the accompanying drawings.

[0163] Please refer to Figure 4, which is a schematic diagram of a scenario provided in an embodiment of the present application. Figure 4 illustrates the initiator, other participants, and coordinator. Other participants refer to participants in distributed learning other than the initiator. Any of the initiator, other participants, and coordinator can be implemented through a network element. There can be one or more other participants, without limitation.

[0164] Figure 4 uses the example of an initiator, network element A, with other participants being network elements B and C, and a coordinator, network element D. The initiator (e.g., network element A) initiates the model inference or distributed learning process. The coordinator assists in the model inference or distributed learning process. The other participants participate in the model inference or distributed process. The implementation of network elements can be referenced in the previous discussion of network elements, and any repetition is omitted.

[0165] The 3GPP standard defines an intelligent network architecture based on NWDAF. This intelligent network architecture aims to collect massive amounts of information from the network and utilize big data and artificial intelligence technologies (such as vertical federated learning) to leverage this data. This output provides valuable information to assist operators in policy formulation and network resource adjustments, thereby improving user experience and reducing network load.

[0166] Please refer to Figure 5, which is a schematic diagram of an intelligent network architecture based on NWDAF provided in an embodiment of the present application, and can also be a schematic diagram of another scenario provided in an embodiment of the present application. Figure 5 illustrates NWDAF, vertical federated learning server (VFL Server), vertical federated learning support function (VFLSF), AF, network function (NF), supervision and management (OAM) (also known as network management), UE and RAN. NF refers to any core network element, such as network exposure function (NEF), network repository function (NRF), network data analytics function (NWDAF), policy control function (PCF), access and mobility management function (AMF), session management function (SMF), operation, or user plane function (UPF), etc.

[0167] Optionally, OAM is deployed on the operator side. The VFL Server can be deployed on the core network side, in the data network (DN) or mobile edge computing (MEC), or the VFL Server can be a third-party server, and there is no specific limitation on this. For example, the VFLSF can be a newly added core network element or access network element, or can be deployed in an existing core network element, for example, deployed in an NRF, NEF, NWDAF or AF network element, and there is no specific limitation on this. Figure 5 is an example of VFLSF and VFL Server being deployed as separate network elements on the core network side, and there is no actual limitation on the implementation of the VFL Server. The implementation of VFLSF can also refer to the content of the implementation of VFL Server, which will not be listed here.

[0168] Any one of the NF, UE or RAN involved in Figure 5 can be used as an example of network element A, network element B or network element C involved in Figure 4, and the NRF, NEF or VFLSF involved in Figure 5 can be used as an example of network element D involved in Figure 4.

[0169] NWDAF has functions such as data collection, model training, data analysis, and model reasoning. It can be used to collect relevant data from other network elements, third-party servers, terminal devices, or network management systems, perform data analysis or model training based on the relevant data, and provide data analysis results to network elements, third-party service servers, terminal devices, or network management systems, or provide trained models to other data analysis network elements. NWDAF can be divided into analytics logical function (AnLF) and model training logical function (MTLF) according to its function. AnLF is a logical function in NWDAF that is used to perform model reasoning, derive analysis results (i.e., derive statistical or predictive analysis results based on the analysis consumer's request), and open analysis results.

[0170] MTLF is a logical function in NWDAF that is used to train models and open training services (for example, providing trained models). An NWDAF may contain only AnLF, only MTLF, or both AnLF and MTLF.

[0171] Please refer to Table 2 below for the analysis results provided by NWDAF and the data that NWDAF needs to collect to provide the corresponding analysis results.

[0172] Table 2

[0173] As shown in Table 2 above, NWDAF can provide service experience analysis results, or it can be described as NWDAF implementing a service experience analysis task. To provide this analysis result, NWDAF needs to collect service-related data such as service identification and service experience from AF / UE, as well as data such as signal reception power and signal reception quality from OAM. NWDAF can train an AI model based on the collected data and then obtain inferred analysis results based on the AI ​​model, such as a predicted service experience for a certain time period in the future. NWDAF can provide network element load analysis results, or it can be described as NWDAF implementing a network element load analysis task. To provide this analysis result, NWDAF needs to collect information related to network element resource usage from NRF, as well as information such as UE speed / direction, and data such as UE path, destination, and arrival time from third-party applications. NWDAF can train an AI model based on the collected data and obtain inferred analysis results based on the AI ​​model, such as a predicted network element load for a certain time period in the future. NWDAF can also provide MEC service experience analysis, or can be described as implementing MEC service experience tasks. In order to provide the analysis results, NWDAF needs to collect UE identification, UE location, application identification, application location and uplink / downlink transmission performance data from AF or UE, and train AI models based on these data, and obtain inference analysis results based on the AI ​​model, such as the predicted MEC service experience analysis for a certain time period in the future.

[0174] The AF is used to convey the requirements of the application side to the network side, such as quality of service (QoS) requirements or user status event subscriptions, etc. The AF can be a third-party functional entity or an application server deployed by the operator.

[0175] VFLSF is responsible for the registration and discovery of participants in vertical federated learning. VFLSF can be deployed independently or in network elements such as NRF, NEF, NWDAF, or AF.

[0176] For example, the VFL Server can act as a coordinator, responsible for maintaining the vertical federation process, authorizing the admission and removal of federated learning members, etc. Optionally, the VFL Server is responsible for tasks such as encryption key distribution and decryption of intermediate information.

[0177] NEF is used to provide the framework, authentication, and interface related to network capability exposure, and to transmit information between 5G system network functions and other network functions.

[0178] NRF is used to provide registration and discovery capabilities for network elements in the network.

[0179] PCF is mainly responsible for the generation and update of UE access strategy and QoS flow control strategy.

[0180] The AMF is mainly responsible for user access and mobility management functions, including user registration, reachability, mobility management, N1 / N2 interface signaling transmission, access authentication and authorization, etc.

[0181] SMF is used to manage the creation, modification, and release of user protocol data unit (PDU) sessions, as well as the allocation and management of IP addresses, the selection and control of UPFs, etc.

[0182] OAM is used to provide system or network fault indication, performance monitoring, security management, diagnosis, configuration and user configuration.

[0183] As a network function element of the 5G core network, UPF undertakes core network processing functions such as data traffic processing and routing forwarding.

[0184] In Figure 5 , Nnef, Nnrf, Nnwdaf, Naf, Npcf, Nvflsf, Namf, Nsmf, and Nvfl-server are service-oriented interfaces provided by the NEF, NRF, NWDAF, AF, PCF, VFLSF, AMF, SMF, and VFL-Server, respectively, for invoking corresponding service-oriented operations. N2 in Figure 5 is the communication interface between the RAN and the core network control plane (e.g., AMF). N3 is the communication interface between the RAN and the UPF, used to transmit user data. N4 is the communication interface between the SMF and the UPF, used for policy configuration, etc., for the UPF.

[0185] Please refer to Figure 6, which is a schematic diagram of a scenario provided in an embodiment of the present application. Figure 6 uses network element A as the AF, network element B / network element C as the NF, and network element D as the VFL Server as an example. In practice, the specific implementation of network elements A, B, C, and D is not limited. In this case, the AF can serve as the initiator of model reasoning or distributed learning, the NF can serve as another participant in model reasoning or distributed learning, and the VFL Server can serve as the coordinator of model reasoning or distributed learning.

[0186] The communication method provided in the embodiments of the present application is introduced below with reference to the accompanying drawings.

[0187] In the various embodiments of the present application, the steps indicated by dotted lines are optional steps. In addition, any network element or a third network element in the at least one network element involved in the various embodiments of the present application is, for example, the master participant or the slave participant involved in FIG3 , the other participants, network element B, or network element C involved in FIG4 , or any network element involved in FIG5 , or network element B, network element C, or NWDAF involved in FIG6 , and the coordinator is, for example, the coordinator involved in FIG3 , or the coordinator or fourth network element involved in FIG4 . The role of the second network element involved in each embodiment of the present application may be a participant. In this case, the second network element is, for example, the master participant or slave participant involved in Figure 2, the initiator or network element A involved in Figure 4, any network element involved in Figure 6, or the network element or AF involved in Figure 6; or the role of the second network element may be a coordinator. In this case, the second network element is, for example, the coordinator involved in Figure 3, the coordinator or network element D involved in Figure 4, the VFL Server involved in Figure 5, or the network element D or VFL Server involved in Figure 6. These network elements may have other names or the names of such network elements may change as the standard continues to evolve, and this is not limited.

[0188] Figure 7 illustrates a communication method provided by an embodiment of the present application. This embodiment uses multiple network elements participating in model reasoning as an example, where these multiple network elements include a first network element. Optionally, the multiple network elements may also include a second network element and / or a third network element, and the number of these multiple network elements is not specifically limited. The following describes the various steps involved in Figure 7.

[0189] S701: A second network element sends first information. This embodiment of the present application takes at least one network element receiving first information from a second network element as an example. The first information indicates a time for determining inference data.

[0190] The at least one network element belongs to some or all of the multiple network elements. When the second network element is a coordinator of distributed reasoning, the at least one network element may be all of the multiple network elements. When the second network element is a participant in distributed reasoning, the at least one network element may be a network element in the multiple network elements other than the first network element.

[0191] In one possible implementation, the second network element may broadcast first information carrying distributed reasoning information, such as a distributed reasoning ID and / or IDs of some or all of the multiple models. In this manner, at least one network element may determine whether to participate in the distributed reasoning based on the distributed reasoning information (e.g., the distributed reasoning ID and / or IDs of some or all of the models).

[0192] Multiple models are associated with a distributed inference ID. Multiple models can be considered models participating in distributed inference, and multiple models correspond to multiple network elements. For example, multiple models correspond one-to-one with multiple network elements, such as each of the multiple network elements deploying one of the multiple models, or each of the multiple network elements using one of the multiple models for inference.

[0193] In another implementation method, client A sends a first message to client B and client C respectively. The first message carries a distributed reasoning ID (that is, the ID is used to identify the joint model of client A / B / C, which can also be called a global model ID). Client B / C determines which model to use to perform distributed reasoning based on the distributed reasoning ID. For example, client B / C locally stores the correspondence between the distributed reasoning ID and the local model to determine which model to use to perform distributed reasoning.

[0194] In another possible implementation, since the second network element can clearly identify the various participants in the distributed reasoning process, the second network element can send the first information to at least one network element respectively, and the first information sent to one network element (such as the first network element) of the at least one network element notifies the first network element to determine the time of reasoning data. The first information sent to the first network element can be referred to as the first information corresponding to the first network element. Under this implementation, optionally, the first information corresponding to the first network element may include the ID and / or distributed reasoning ID of the model of the first network element (such as the first model). The first model is one of the multiple models discussed above.

[0195] For example, the second network element is client A, and at least one network element includes client B and client C. Client B and client C both participate in the distributed training process and each has its own locally trained model. For example, client A corresponds to model 1, client B corresponds to model 2, and client C corresponds to model 3. Then client A can send a first message to client B, where the first message carries the identifier of model 2 corresponding to client B, and send a first message to client C, where the first message carries the identifier of model 3 corresponding to client C. In this way, after client B receives the first message, it can clearly use model 2 to participate in distributed reasoning, and after client C receives the first message, it can clearly use model 3 to participate in distributed reasoning.

[0196] The role of the second network element in distributed reasoning is different, so the relationship between at least one network element and multiple models is also different, which are explained below.

[0197] 1. The second network element is the coordinator of distributed reasoning.

[0198] In this case, the multiple network elements do not include the second network element, and the at least one network element is all of the multiple network elements, that is, the at least one network element is a network element that participates in model inference. Accordingly, multiple models can correspond to at least one network element. Multiple models can correspond one-to-one with at least one network element, or each network element in at least one network element can correspond to one of the multiple models. For example, each model in the multiple models is deployed in one of the at least one network elements. Alternatively, multiple models can correspond one-to-many with at least one network element, or each network element in at least one network element can correspond to at least two of the multiple models. For example, at least two of the multiple models are deployed in one of the at least one network element.

[0199] 2. The second network element is a participant in distributed reasoning.

[0200] In this case, the multiple network elements include the second network element, and the at least one network element is a network element other than the second network element in the multiple network elements. For example, the models other than the model of the second network element (e.g., the second model) in the multiple models can correspond one-to-one with the at least one network element, or can be described as each network element in the at least one network element corresponding to one model other than the second model in the multiple models. For example, each model other than the second model in the multiple models is deployed in one of the at least one network elements. Alternatively, the models other than the second model in the multiple models can correspond one-to-many with the at least one network element.

[0201] The following describes the content of the first information corresponding to the first network element, indicating the time for determining inference data, using the first information as an example. Inference data refers to inference data used by the first network element to perform model inference based on the first model. The first model is associated with one or more models other than the first model in the multiple models. In other words, the first model and the one or more models all belong to the multiple models discussed above, or the multiple models include the first model and the one or more models.

[0202] The first information corresponding to the first network element includes at least one of the following information A1 to A5 to indicate the time for determining the inference data, which are introduced below. It can be understood that at least one of the following information A1 to A5 can be regarded as an implementation method of the first information indicating the time for determining the inference data.

[0203] A1. The first information includes first timestamp information. The first timestamp information is used to determine the time of the inference data.

[0204] The first timestamp may be in the form of an integer or a floating point number, etc., without limitation. The first timestamp is used to determine or indicate a first moment related to the first data (or referred to as a reference moment, etc.). The first moment refers to a moment related to the inference data. The first timestamp may include one or more timestamps, without specific limitation. In the case where the first timestamp includes multiple timestamps, these multiple timestamps may be periodically distributed or non-periodically distributed, without limitation.

[0205] For example, the first moment includes a relevant moment of collecting (or collecting) reasoning data, such as the moment of starting to collect reasoning data, or the moment of ending to collect reasoning data, or the moment in the middle of collecting reasoning data, etc., without limitation. In the embodiment of the present application, the reasoning data of the first network element is taken as the first data as an example. The first data corresponding to the first network element (or the first model) can be regarded as an example of reasoning data.

[0206] For the first data, the first moment includes, for example, the first moment when the first network element collects the first data, for example, the moment when the first network element collects or detects the first data from the data generated by itself, or the moment when the first network element collects the first data from the data generated by other network elements. Alternatively, the first moment includes a moment related to the generation (or production) of the first data, for example, the moment when the generation of inference data begins, or the end moment when the generation of inference data ends, or the intermediate moment when the generation of inference data is generated, etc., without limitation. Optionally, the first data corresponding to the first network element (or the first model) may be an example of inference data. For the first data, the first moment includes, for example, the moment when the first network element generates the first data, or the moment when other network elements generate the first data.

[0207] The following describes the content of the first timestamp information with examples in conjunction with B1 to B3.

[0208] B1. The information of the first timestamp includes information of the first moment. Alternatively, it can be described as the first information indicating a reference moment, or it can be described as the information of the first timestamp indicating the first moment.

[0209] Under B1, correspondingly, at least one network element may determine inference data based on the first moment. For example, the inference data meets the requirements of the first moment. For example, the inference data includes first data, and the first data may be determined based on the first moment.

[0210] For example, if the first moment indicates that data collection starts at 13:00 on January 1, 2024, then the first network element can determine that the data collection starts at 13:00 on January 1, 2024 for model inference, which means that the time when the first data collection starts is 13:00 on January 1, 2024. For another example, if the first moment includes [19:40, 20:00, 20:10], and the first moment indicates the start time of data collection, then the first network element can use data with the start time of data collection being 19:40, 20:00, or 20:10 as data for model inference, which means that the time when the first data collection starts is 19:40, 20:00, or 20:10.

[0211] B2. The information of the first timestamp includes information of a time deviation (or time offset, etc.). Alternatively, it can be described as the first information indicating the time offset, or it can be described as the information of the first timestamp indicating the time offset.

[0212] Time offset refers to the deviation in the time allowed for collecting or generating inference data, or can be described as the acceptable data collection time deviation from the first timestamp. The time offset and the first moment can be used to determine the time range of the first data.

[0213] The meaning of the first moment is different, so the specific meaning of the time offset is also different. The following examples are given to illustrate.

[0214] For example, if the first moment is the moment when the inference data collection starts, then the time offset represents the maximum deviation of the moment when the inference data collection starts relative to the first moment. Alternatively, if the first moment is the moment when the inference data collection ends, then the time offset represents the maximum deviation of the moment when the inference data collection ends relative to the first moment. Alternatively, if the first moment is the moment when the inference data generation starts, then the time offset represents the maximum deviation of the moment when the inference data generation starts relative to the first moment. Alternatively, if the first moment is the moment when the inference data generation ends, then the time offset represents the maximum deviation of the moment when the inference data generation ends relative to the first moment.

[0215] Under B2, the first network element determines the inference data to use based on the first time instant and the time offset. For example, the inference data meets the requirements of the first time instant and the time offset. Under B2, the first time instant can be preconfigured or predefined in the first network element, or defaulted to the current time instant, or determined by negotiation between the first network element and the second network element, without limitation. For example, the inference data includes the first data, and the first data can be determined based on the first time instant and the time offset.

[0216] For example, the time offset is 10 minutes (min), and the first moment indicates that the time when data collection starts is 13:10 on January 1, 2024. Then the first network element can use the data between 13:00 on January 1, 2024 and 13:20 on January 1, 2024 as the inference data of the first model, which means that the time when the first data collection starts is between 13:00 on January 1, 2024 and 13:20 on January 1, 2024.

[0217] B3. The first timestamp information includes information about the first moment and time offset information. Alternatively, it can be described as the first information indicating the first moment and time offset, or it can be described as the first timestamp information indicating the first moment and time deviation.

[0218] Under B3, the first network element accordingly determines the inference data based on the first moment and the time offset. For example, the inference data meets the requirements of the first moment and the time offset. For example, the inference data includes the first data, and the first data can be determined based on the first moment and the time offset.

[0219] For example, the first moment indicates that the end time of data collection is 19:40 every day, and the time offset is 10 minutes. Then the first network element uses the data when the data collection ends between 19:30 and 19:50 every day as the inference data of the first model, which means that the end time of collecting the first data is between 19:30 and 19:50 every day.

[0220] A2. The first information includes information of a first time window (data time window(s)).

[0221] The first time window represents one or more time windows used to infer data association. The information of the first time window includes at least one of the length of the first time window, the start time, the end time, or the first period. The length of the first time window is, for example, the time interval between the start time of the first time window and the end time of the first time window. In the case where the first time window appears periodically, the first period represents the period in which the first time window appears, or can be understood as the time interval between two periodically appearing first time windows. For example, if the first time window is 19:00-20:00 and the first period is 24 hours, then it means that the first time window corresponds to 19:00-20:00 every day.

[0222] Optionally, the information of the first time window includes the length of the first time window, and the start time or end time of the first time window may be pre-configured or pre-defined in the first network element, or default to the current time, or determined by negotiation between the first network element and the second network element, and there is no limitation on this. Alternatively, the information of the first time window includes the start time of the first time window, and the end time or length of the first time window may be pre-configured or pre-defined in the first network element, or default to the current time, or determined by negotiation between the first network element and the second network element, and there is no limitation on this.

[0223] Alternatively, the information of the first time window includes the end time of the first time window, and the start time or length of the first time window can be pre-configured or pre-defined in the first network element, or default to the current time, or determined by negotiation between the first network element and the second network element, and there is no limitation on this. Alternatively, the information of the first time window includes the start time and the first period of the first time window, which means that model inference can be performed once for every interval of the first period. For example, the start time of the first time window is 19:40, and the start time indicates the time when data collection starts, and the first period is 10 minutes, which means that the data collected at 19:40 is used as inference data, and the data collected at 19:50, 20:00, and 20:10... can also be used as inference data. Alternatively, the information of the first time window includes the end time and the first period of the first time window, which means that model inference can be performed once for every interval of the first period.

[0224] The first time window can be a time window related to data collection. For example, the first time window indicates that the collection or generation of inference data begins within the time range indicated by the time window. Alternatively, the first time window indicates that the collection or generation of inference data ends within the time range indicated by the time window. Alternatively, both the time when the collection or generation of inference data begins and the time when the collection or generation of inference data ends fall within the time range indicated by the first time window. For example, the start time (or starting time) of the first time window is the time when the collection of inference data begins, and the end time (or ending time) of the first time window is the time when the collection of inference data ends. Alternatively, the first time window can be a time window related to data generation, for example, the start time of the first time window is the time when the generation of inference data begins, and the end time of the first time window is the time when the generation of inference data ends. Alternatively, the start time of the first time window is the time when the generation of inference data begins, and the end time of the first time window is the time when the collection of inference data ends, etc. It suffices for the second network element and the first network element to reach a consensus on the first time window. There may be various interpretations of the first time window, and this is not specifically limited.

[0225] For example, the information of the first time window represents the time window related to the collected data, and the first time window is [19:40, 20:00]. Then the first network element can use the data with the start time of collection at 19:40 and the end time of collection at 20:00 as the inference data of the first model.

[0226] For another example, the information of the first time window represents the time window related to the collected data, and the information of the first time window includes [19:40, 20:00], [20:10, 20:15] and [20:30, 20:40], then it means that the first network element will use the data collected at [19:40, 20:00] as inference data, the data collected at [20:10, 20:15] as inference data, and the data collected at [20:30, 20:40] as inference data of the first model.

[0227] In one possible design, the first information includes information about the first timestamp and information about the first time window. The information about the first timestamp can refer to the content of the first timestamp information discussed in A1 above, and the information about the first time window can refer to the content of the first time window information discussed in A2 above. Any repetitions are not listed here. For example, the first information can include the start time, first period, and time offset of the first time window. Alternatively, the first information can include the end time, first period, and time offset of the first time window.

[0228] For example, the start time of the first time window is 19:40, the end time of the first time window is 19:50, the first time window represents the time range of the time when the data is collected, the first cycle is 10 minutes, and the time offset is 1 minute. Then the first network element can

[0229] The data collected between 19:51 and 19:48 is used as inference data, the data collected between 19:48 and 20:02 is used as inference data, and so on.

[0230] For another example, the information of the first timestamp indicates that the first moment to start generating inference data is 18:00, and the first time window is [19:40, 20:00]. This means that the first network element will start generating inference data at 18:00, and the data collected at [19:40, 20:00] will be used as the inference data of the first model.

[0231] A3. The first information includes information about the time when the inference was performed based on the model. This can be described as the first information indicating the time used to determine the inference data including: The first information includes information about the time when the inference was performed based on the model. Alternatively, this can be described as information about the time when the inference was performed based on the model being used to determine the inference data.

[0232] The time of performing reasoning based on the model is, for example, the time of starting reasoning based on the model, or the time of ending reasoning based on the model, or any time of performing reasoning based on the model, which is not specifically limited.

[0233] Illustratively, the first network element preconfigures or predefines a first relationship between the time when the model-based reasoning is performed and the time of the reasoning data. Thus, the first network element can determine the time of the reasoning data based on the time when the model-based reasoning is performed and the first relationship. The following model uses the first model as an example to illustrate possible content of the first relationship.

[0234] For example, the time of inference data is the time when inference data collection begins, and the first relationship indicates that the difference between the time when model-based inference begins and the time when inference data collection begins is the first duration. The first duration is, for example, a number greater than 0, such as 5, 10, or 30 minutes, and is not specifically limited to this. For example, if model-based inference begins at 12:00 and the first duration is 30 minutes, then the time when inference data collection begins is 11:30.

[0235] A4. The first information includes information about the first condition. This can be described as the first information indicating a time for determining the inference data including: the first information includes information about the first condition. Alternatively, this can be described as the first condition being used to determine the time for determining the inference data.

[0236] The first condition may also be referred to as an inference trigger condition. It indicates the conditions under which the first network element triggers inference execution, that is, under which conditions (such as certain events) the first network element triggers model inference. When the at least one network element includes more than the first network element, the inference trigger conditions corresponding to the at least one network element may be the same or different, and this is not specifically limited.

[0237] For example, the first condition includes monitoring that the terminal device moves to a specific cell, or monitoring that the terminal device switches the network mode (for example, the terminal device switches from 4G access to 5G access), or monitoring that the location of the terminal device moves to a specific location, etc., and there is no specific limitation on this. In the case where the first condition includes monitoring that the terminal device moves to a specific cell and / or monitoring that the location of the terminal device moves to a specific location, the first network element may be directly monitoring the location of the terminal device, or the first network element may be subscribing to or obtaining the location of the terminal device from other network elements (such as NWDAF), and there is no specific limitation on this.

[0238] A5. The first information includes first indication information. This can be described as the first information indicating a time for determining the inference data including: the first information includes the first indication information. Alternatively, this can be described as the first indication information being used to determine the time for determining the inference data.

[0239] The first indication information indicates to start reasoning based on the model, or the first indication information indicates to execute model reasoning immediately. The first indication information indicating to execute model reasoning immediately can be understood as that, after the first network element receives the first indication information, it immediately starts to collect reasoning data and performs model reasoning based on the collected reasoning data; or, if the first network element determines that the reasoning data has been collected after receiving the first indication information, the first network element immediately executes model reasoning based on the collected reasoning data. The first indication information can also be called a model inference indication. For example, the first indication information is used to indicate to execute reasoning based on at least one model among one or more models. If at least one model includes the first model, then the first indication is used to instruct the first model to execute reasoning. The following is introduced by taking at least one model including the first model as an example.

[0240] For example, the first indication information instructs the first network element to perform reasoning based on a model (such as the first model). The first network element may then collect reasoning data and perform reasoning based on the model. Alternatively, the first indication information is used to instruct the immediate execution of reasoning based on some or all of the multiple models. After receiving the first indication information, the first network element may use the most recently collected data as reasoning data and perform reasoning based on the model.

[0241] In another possible implementation, if the first information is the information shown in A1 or A2 above, the first information may further include at least one of information about the time when model-based reasoning is started, information about the first condition, or first indication information. In this case, at least one of the information about the time when model-based reasoning is started, information about the first condition, or first indication information is not used to determine the time of the inference data.

[0242] Optionally, the first information may further indicate a mode for sending the inference result. For example, the mode for sending the inference result indicates that the inference result is calculated and reported each time after the inference data is collected. For another example, the mode for sending the inference result indicates whether the inference result is sent periodically or aperiodically. In this way, the first network element may determine the timing of sending the inference result based on the mode for sending the inference result indicated by the first information.

[0243] The content of the first information sent by the second network element to at least one network element can refer to the content of the first information corresponding to the first network element discussed above, and will not be listed one by one here.

[0244] S702. The first network element determines an inference result of the first model based on the first data and the first model.

[0245] After receiving the first information, the at least one network element may determine the inference data according to the first information. Taking the example that the at least one network element includes a first network element and the model of the first network element is a first model, the contents involved in S702 are introduced by way of example.

[0246] After receiving the first information, the first network element can determine the inference data based on the first information. In the embodiment of the present application, the first network element uses the first data as the inference data as an example. In this example, the first network element can determine the input of the first model based on the first data, and input the input of the first model into the first model to obtain the inference result of the first model. The input of the first model can be in the form of a vector, an image, or a matrix, etc., and there is no specific limitation on this. Optionally, the first data is the input of the first model, or the first data is preprocessed to obtain the input of the first model. The content of the preprocessing can refer to the content of the preprocessed data discussed above and will not be listed here.

[0247] For example, the first model includes an output layer and a normalization layer. The output layer in the model can be expressed as Z=X*W+b, where Z represents the output vector of the output layer, X represents the input of the first model, W represents the weight matrix of the first model, b represents the bias, and the normalization layer of the first model is, for example, P i =exp(Z i) / sum(exp(Z)), where exp is the exponential function and Z i is the i-th element of the output vector Z, P i Indicates the classification of the i-th element.

[0248] The input to the first model can be, for example, 0, -135, -5, and 200 indicate that user 1's access type is satellite access, reference signal received power, reference signal received quality, and radio access network throughput, respectively. 1, -80, -10, and 300 indicate that user 2's access type is new air interface access, reference signal received power, reference signal received quality, and radio access network throughput, respectively.

[0249] After the input of the first model passes through the output layer of the first model, an output vector Z can be obtained. After the output vector Z is processed by the normalization layer, the business experience analysis result is output. For example, the business experience analysis result is 4 (that is, the average opinion score of the business experience is 4).

[0250] Optionally, in the case where the first data includes multiple data, the first network element can obtain multiple inference results based on the multiple data and the first model, that is, the inference result of the first model can include multiple inference results.

[0251] The following describes an example of how the first network element determines the first data in conjunction with the implementation shown in C1 or C2.

[0252] C1. The data determined by the first network element according to the first information includes at least one data, and the at least one data is used as the first data.

[0253] C2. The first network element determines, based on the notification from the second network element, to use the first data among the at least one data as the inference data.

[0254] Exemplarily, the data determined by the first network element based on the first information includes at least one data, and the first network element sends information about the at least one data to the second network element. The second network element can determine the first data from the at least one data and send fourth information to the first network element, where the fourth information instructs the first network element to use the first data for inference, or can be described as instructing the first network element to use the first data as inference data. Under this embodiment, since the second network element combines information about the data fed back by at least one network element to determine the inference data used by at least one network element to participate in model inference, this is conducive to more accurate determination of the inference data.

[0255] The information of at least one data item is used to indicate the at least one data item. The information of at least one data item includes information about each data item in part or all of the at least one data item. The at least one data item includes first data. The information of each data item is described below using the information of the first data item as an example. The information of the first data item includes at least one item of information from D1 to D5. The information shown in D1 to D5 is described below.

[0256] D1. Information about the number of the first data.

[0257] The numbering of the first data may be determined by negotiation between the second network element and the first network element, or determined by the first network element itself. For example, the first network element obtains the numbering by processing the time information of the first data, such as performing a hash calculation on the time information of the first data. Alternatively, the first network element sequentially numbers the collected data in the order of the time when the data was collected, or in the order of the time when the data was generated, to obtain the numbering of the first data. The time of the data here indicates the time when the data was actually collected or generated, for example, a timestamp indicating the actual collection or generation of the data, or a time window when the first data was actually collected or generated.

[0258] For example, the first network element collects the average throughput rate of a certain UE three times within the first time window indicated by the first information. The numbers of the three collected average throughput rates of UE1 may be, for example, 001, 002, and 003, respectively.

[0259] Under D1, by analogy, at least one network element can provide feedback to the second network element regarding the number of at least one data item determined based on the first information. The second network element can then obtain the number of at least one data item corresponding to each of the multiple models. Thus, the second network element can determine the first data item from the at least one data item corresponding to the first network element based on the number of at least one data item corresponding to each of the multiple models. Optionally, the fourth information can include information regarding the number of the first data item, which is used to instruct the first network element to use the first data item as inference data.

[0260] The following describes an example of a method in which the second network element determines the first data from at least one data corresponding to the first network element in the case of D1.

[0261] E1. The second network element may use the smallest-numbered data among the at least one data corresponding to each model in the multiple models as the inference data corresponding to each network element. In this case, the first data may be the smallest-numbered data among the at least one data corresponding to the first network element. If the numbers can be determined in chronological order of collection or generation, selecting the smallest-numbered data as the inference data can maximize the proximity of the collection or generation times of the inference data participating in the distributed inference, thereby improving the accuracy of the model inference.

[0262] For example, taking the example of a model participating in distributed reasoning including a first model and a second model, the at least one data corresponding to the first network element (or the first model) includes two data, which are numbered 001 and 002, and the at least one data corresponding to the second network element (or the second model) includes four data, which are numbered 003, 004, 005, and 006. The second network element may determine the data numbered 001 as the first data, and determine the data numbered 003 as the inference data corresponding to the second network element (hereinafter referred to as the second data).

[0263] E2. The second network element may use the data with the largest number among the at least one data corresponding to each network element in the at least one network element as the inference data corresponding to the network element. In this case, the first data may be the data with the largest number among the at least one data corresponding to the first network element. If the numbers can be determined according to the chronological order of collection or generation, selecting the data with the largest number as the inference data can maximize the proximity of the collection or generation times of the inference data participating in the distributed inference, thereby improving the accuracy of the model inference.

[0264] For example, taking the example of a model participating in distributed reasoning including a first model and a third model, the first model corresponding to a first network element, and the third model corresponding to a network element other than the first network element (such as a third network element) in at least one network element, the N data (i.e., at least one data) corresponding to the first network element (or the first model) include 3 data, and the 3 data are numbered 001, 002, and 003, respectively, and the at least one data corresponding to the third network element (or the second model) includes 3 data, and the 3 data are numbered 001, 002, and 003, respectively. The second network element may determine the data numbered 003 in the at least one data corresponding to the first network element as the first data, and determine the data numbered 003 in the at least one data corresponding to the third network element as the inference data corresponding to the third network element.

[0265] E3. The second network element may determine as first data the data whose number difference between the at least one data corresponding to the first network element and the inference data of one or more models is less than or equal to the first threshold. Under E3, the difference between the number of the first data and the number of the inference data of one or more models is less than or equal to the first threshold. The content of one or more models can refer to the content of the one or more models discussed above and will not be listed here. The first threshold may be pre-configured or pre-defined in the second network element, or determined by negotiation between the second network element and at least one network element, and this is not limited. The second network element may have already clearly defined the inference data corresponding to one or more models. In the case where the number can be generated according to the time of collection or generation, or determined according to the order of the time of collection or generation, data with relatively small number difference are selected as inference data, which can make the time of collection or generation of inference data participating in distributed inference as close as possible, and can also improve the accuracy of model inference.

[0266] For example, taking the case where multiple models include a first model and a second model, and the first threshold is 4, the N data (i.e., at least one data) corresponding to the first network element (or the first model) include 3 data, and these 2 data are numbered 001, 002, and 003, respectively, and the at least one data corresponding to the second network element (or the second model) includes 4 data, and these 4 data are numbered 004, 005, 006, and 007, respectively. The second network element uses the data numbered 007 as the inference data of the second model. Since the difference between the numbers of the data numbered 003 and the inference data of the second model is 4, the second network element can determine the data numbered 003 as the first data.

[0267] D2. Data characteristic information: The data characteristic information indicates the characteristics (or attributes) of the first data.

[0268] The data feature information may include feature information, such as at least one of location, average throughput, or average packet delay. The feature information may include, for example, a feature identifier, which indicates the feature.

[0269] The identification of a feature, which may be referred to as a feature identification, may be an identification assigned to the feature (which may also be referred to as a number, serial number, or index, etc.). In one possible implementation, the identification of a feature may be understood as an anonymous feature, and the participants (such as the first network element and other participants) negotiate the meaning of the feature identification in advance, so that they can understand each other's meaning of the identification of the feature registered by the other party. Optionally, the coordinator cannot know which feature the feature identification specifically represents through the feature identification, so as to avoid leaking the features supported by each participant to other parties other than the participants in the distributed learning. In other words, the identification of the feature can be understood by the participants in the distributed learning model, but not by the second network element, thus ensuring the security of the feature. In another possible implementation, the identification of the feature may also be non-anonymous, and the participants do not need to negotiate the identification of the feature in advance, and the embodiments of the present application do not need to limit this.

[0270] For example, the feature identifier is 1234, indicating that the feature is average throughput; or the feature identifier is 1235, indicating that the feature is average packet delay.

[0271] Under D2, by analogy, at least one network element can respectively feedback data feature information of at least one data determined based on the first information to the second network element, and the second network element can obtain the data feature information corresponding to each of the multiple models. In this way, the second network element can determine the first data from the at least one data corresponding to the first network element based on the features corresponding to each of the multiple models. Alternatively, the second network element can determine the first data that meets the distributed reasoning requirements from the at least one data corresponding to the first network element based on the requirements of distributed reasoning.

[0272] For example, the second network element determines that the feature used by the first network element to participate in distributed reasoning is the average throughput, and the second network element determines that the feature corresponding to a certain data in at least one data is the average throughput, then the second network element can determine the data as the first data.

[0273] Optionally, the fourth information may include data feature information corresponding to the first data and / or information on the number of the first data, and the data feature information and / or information on the number of the first data are used to instruct the first network element to use the first data as inference data.

[0274] D3. First time information: The first time information indicates the time when the first data is actually collected or generated.

[0275] The first time information may include, for example, information about a second timestamp and / or information about a second time window. The second timestamp information may indicate the actual time when the inference data was collected or generated, and may include, for example, a second time. The second time may be, for example, the actual time when the collection of the first data actually began, the actual time when the generation of the first data actually began, the actual time when the collection of the first data actually ended, or the actual time when the generation of the first data actually ended.

[0276] The second time window is the time window during which the first data is actually collected or generated. For example, the second time window includes at least one of the start time, end time, length, or second period of the second time window. The second period is the period during which the second time window actually occurs. For example, the start time of the second time window is the actual start time of collecting the first data, and the end time of the second time window is the actual end time of collecting the first data. Alternatively, the start time of the second time window is the actual start time of generating the first data, and the end time of the second time window is the actual end time of generating the first data, etc. The details of the second time window can be referred to the details of the first time window discussed above and will not be listed here.

[0277] Under D3, by analogy, at least one network element can respectively feedback the first time information of at least one data determined based on the first information to the second network element, and the second network element can obtain the first time information corresponding to each model in the models participating in distributed reasoning. In this way, the second network element can determine the first data based on the first time information corresponding to each model in the models participating in distributed reasoning and from at least one data corresponding to the first network element.

[0278] In one possible implementation, the second network element may determine as first data the data whose time difference between at least one data corresponding to the first network element and the inference data corresponding to one or more models is less than or equal to a second threshold. In this case, the time difference between the actual collection or generation time of the first data and the inference data corresponding to one or more models is less than or equal to the second threshold. The manner in which the second network element obtains the second threshold can refer to the content of the second network element obtaining the first threshold discussed above and will not be listed here.

[0279] In the case where the first time information corresponding to the first network element indicates a second time window, the time difference between the first data and the inference data corresponding to one or more models can be the difference between the start times of the two second time windows corresponding to the first data and the inference data corresponding to one or more models, or the difference between the end times of the two second time windows, or the sum of the difference between the start times of the two second time windows and the difference between the end times of the two second time windows, etc., and there is no specific limitation on this.

[0280] For example, the models participating in distributed reasoning include a first model and a second model, the reasoning data corresponding to the first model includes data 1, the at least one data corresponding to the second model includes data 2 and data 3, and the second threshold is 1. The second time information corresponding to the first model indicates that the time when data 1 collection starts is 18:00. The second time information corresponding to the second model indicates that the time when data 2 collection starts is 18:01, and the time when data 3 collection starts is 18:02. Because the difference between the time when data 2 collection starts and the time when data 1 collection starts is 1, the second network element can determine data 2 as the first data.

[0281] In another possible implementation, the second network element may determine as the first data the data with the smallest sum of the time of the inference data corresponding to one or more models among the at least one data corresponding to the first network element. In this case, the sum of the time of the actual collection or generation of the first data and the time of the inference data corresponding to the one or more models is the smallest. The meaning of the one or more models can refer to the content of the one or more models involved in D1 above and will not be listed here.

[0282] Optionally, the fourth information may include first time information corresponding to the first data and / or information on the number of the first data, and the first time information and / or information on the number of the first data are used to instruct the first network element to use the first data as inference data.

[0283] D4. First location information: The first location information indicates the location where the first data is collected.

[0284] The location here can be a specific geographic location, a geographic area, a cell, a combination of multiple cells, a tracking area, or a combination of multiple tracking areas. For example, if the first data is collected within a certain area (such as an area of ​​interest (AOI)), then the location of the first data is that area. Alternatively, if the first data is collected within a certain location, then that location is the location of the first data. The AOI is, for example, a combination of one or more tracking areas (TAs) or a combination of one or more cells.

[0285] Under D4, by analogy, at least one network element can respectively feedback to the second network element the first position information of at least one data determined based on the first information, and the second network element can obtain the first position information corresponding to each model in the multiple models. In this way, the second network element can determine the first data based on the first position information corresponding to each model in the multiple models and from at least one data corresponding to the first network element.

[0286] In one possible implementation, the second network element may determine as first data data whose distance between the location of the inference data corresponding to one or more models and the location of at least one data corresponding to the first network element is less than or equal to a third threshold. In this case, the distance between the location of the first data and the location of the inference data corresponding to the one or more models is less than or equal to the third threshold. The third threshold may be preconfigured or predefined in the second network element, or determined by negotiation between at least one network element and the second network element, and this is not limited.

[0287] D5. Access type information. The access type information indicates the access type corresponding to the first data. For example, the access type corresponding to the first data is 4G or 5G.

[0288] Under D5, by analogy, at least one network element can respectively feedback to the second network element the access type information of at least one data determined based on the first information, and the second network element can obtain the access type information corresponding to each model in the models participating in distributed reasoning. In this way, the second network element can determine the first data from at least one data corresponding to the first network element based on the access type information corresponding to each model in the models participating in distributed reasoning.

[0289] Exemplarily, the second network element may determine, as the first data, data of the same access type as the inference data corresponding to the one or more models in the at least one data corresponding to the first network element. In this case, the location of the first data is the same as the access type corresponding to the inference data corresponding to the one or more models.

[0290] If the information of the at least one data corresponding to the first network element includes at least two items of information from D1 to D5, the second network element may also determine the first data from the at least one data corresponding to the first network element based on one of the at least two items of information. Alternatively, the second network element may determine the first data from the at least one data corresponding to the first network element based on the contents corresponding to the at least two items of information.

[0291] For example, if the information of at least one data corresponding to the first network element includes the above-mentioned D1 (i.e., the data number), D3 (i.e., the first time information) and D5 (i.e., the access type information), then the second network element can determine the data in the at least one data corresponding to the first network element, whose time difference with the inference data corresponding to one or more models is less than or equal to the second threshold, as the first data.

[0292] For another example, the information of at least one data corresponding to the first network element includes the above-mentioned D3 (i.e., the first time information), D4 (i.e., the first location information) and D5 (i.e., the access type information). Then the second network element can determine the at least one data corresponding to the first network element, the data whose time difference with the inference data corresponding to one or more models is less than or equal to the first threshold, the difference in location distance is less than or equal to the third threshold, and the access type is the same as the first data.

[0293] For example, multiple network elements include client B and client C. Client B or client C can be used as an example of the first network element. Client B collected 10 pieces of data for a certain feature (such as average throughput), and the timestamp, location, access type, etc. corresponding to each piece of data may be different. Client C collected 15 pieces of data for a certain feature (such as RSRP), and the timestamp, location, access type, etc. corresponding to each piece of data may also be different. The second network element can, for example, determine one piece of data from the 10 pieces of data collected by client B for the average throughput based on the principle of similar timestamps, and determine one piece of data from the 15 pieces of data collected by client C for RSRP, so that the determined data of client B and client C are as close as possible in collection time to improve the accuracy of model inference.

[0294] Of course, the second network element may also determine a piece of data for the model for client B and client C based on other principles (such as location and access type). For example, the other principle is that the initiator determines a piece of data for client B and client C respectively, which has the same location and access type, and the data is collected as close as possible in time. The embodiment of the present application does not specifically limit the principle for the second network element to determine the first data.

[0295] In another possible implementation, the second network element determines that there is no first data meeting the corresponding condition in at least one data corresponding to the first network element, or it can be described as that the first network element does not have data that meets the time for model inference, then the first network element can send a second indication message to at least one network element. The second indication message indicates that the model inference is terminated. Optionally, the second indication message may carry a cause value for model inference termination (model inference termination with cause code), which, for example, indicates that some network elements do not have data that meets the distributed inference.

[0296] S703: The first network element sends second information to the second network element. Accordingly, the second network element receives the second information from the first network element. The second information indicates the inference result of the first model. Alternatively, the second information may include the inference result of the first model.

[0297] After obtaining the first information, the first network element may send the second information to the second network element. If the second network element is also a participant in the distributed reasoning, the first network element may send the second information directly to the second network element, or the first network element may send the second information to the second network element through a coordination direction, without specific limitation.

[0298] By analogy, the second network element can obtain the inference results of the model respectively fed back by at least one network element to obtain the inference result corresponding to at least one network element. Optionally, if the second network element is a participant in distributed inference, the second network element can also determine the second data based on the first information. The second network element determines the inference result of the second model based on the second data and the second model. The content of the second data can be determined by referring to the content of the first data determined above, and the content of the inference result of the second model can also be determined by referring to the content of the inference result of the first model determined above, and they will not be listed one by one here.

[0299] In addition to including information about the inference results of the first model, the second information may optionally further include first time information and / or the order of multiple inference results included in the inference results of the first model. The content of the first time information can refer to the content of the first time information discussed above and is not further detailed here.

[0300] In the case where the second information also includes the first time information, optionally, the second network element can verify whether the inference data of the first model is appropriate based on the first time information. If the error between the time corresponding to the first data and the time indicated by the first information for model inference is greater than or equal to the first error, indicating that the first data (or described as the inference result of the first model) is not appropriate, then the second network element can send a second indication information to one or more second network elements. The content of the second indication information can refer to the content of the second indication information discussed above. If the error between the time corresponding to the first data and the time indicated by the first information for model inference is less than the first error, it indicates that the inference result of the first model is appropriate, or is described as appropriate for the first data. Since the inference result of the first model is further verified based on the first time information, the accuracy of the model inference can be further improved.

[0301] When the second information also indicates the order of multiple inference results included in the inference result of the first model, these multiple inference results correspond to multiple data in the first data, for example, one of the inference results corresponds to one of the multiple data in the first data, or it can be described as an inference result determined based on one of the multiple data in the first data and the first model. The order of these multiple inference results can be, for example, a chronological order corresponding to the time of the inference data corresponding to these multiple inference results, or a chronological order of these multiple inference results can also be generated. The information on the order of multiple inference results includes, for example, the identifiers of multiple inference results (also referred to as inference IDs). The identifiers of multiple inference results can be the timestamps corresponding to the inference data corresponding to these multiple inference results, or a serial number generated according to the chronological order of the generation of the inference results, or a combination of the two, etc., and there is no specific limitation on this.

[0302] For example, the first data includes data 1, data 2 and data 3 collected sequentially, and the multiple inference results corresponding to data 1, data 2 and data 3 are inference result 1, inference result 2 and inference result 3 respectively. Then the order of these multiple inference results can be, for example, inference result 1, inference result 2 and inference result 3, and the information of the order of these multiple inference results can be, for example, 123.

[0303] S704: The second network element determines a first reasoning result based on the reasoning result of the first model. The first reasoning result may also be referred to as a global reasoning result.

[0304] The role of the second network element in distributed reasoning is different, so the process of the second network element determining the first reasoning result is also different. The following describes the different situations.

[0305] F1, in the second network element, is a participant in distributed reasoning.

[0306] The second network element may determine the first inference result based on the inference result corresponding to at least one network element (including the inference result of the first model) and the inference result of the second model. For example, the first inference result may be an aggregated result of the inference result corresponding to at least one network element and the inference result of the second model. Specifically, for example, the first inference result may be a weighted sum of the inference result corresponding to at least one network element and the inference result of the second model, or may be the product of the inference result corresponding to at least one network element and the inference result of the second model. Alternatively, the second network element may use the inference result of at least one network element as input to the model of the second network element (such as the second model) to obtain the first inference result. Alternatively, the second network element may use the inference result of a portion of the at least one network element as input to the model of the second network element, calculate an intermediate inference result, and aggregate the intermediate inference result with the inference result of another portion of the at least one network element to obtain the first inference result. The embodiments of the present application do not limit the specific method for determining the first inference result.

[0307] The following takes at least one network element including a first network element as an example to introduce the content of determining the first reasoning result.

[0308] Exemplarily, the second network element may aggregate the reasoning result of the first model and the reasoning result of the second model to obtain the first reasoning result.

[0309] In the case that the inference result of the second model includes multiple inference results, the second network element may respectively determine the first inference result corresponding to each of the multiple inference results of the second model, that is, obtain multiple first inference results.

[0310] For example, the second network element may calculate a first inference result based on one inference result from the multiple inference results of the second model and an inference result from the multiple inference results of the first model that has the same order as the one inference result, and so on, thereby obtaining multiple first inference results.

[0311] For example, if the inference results of the first model include inference results 1, 2, and 3, and the inference results of the second model include inference results 4, 5, and 6, then the second network element can aggregate inference results 1 and 4 to obtain a first inference result. Furthermore, the second network element can aggregate inference results 2 and 5 to obtain a first inference result. Furthermore, the second network element can aggregate inference results 3 and 6 to obtain a single inference result.

[0312] F2, the second network element, is the coordinator of distributed reasoning.

[0313] The second network element may determine the first inference result based on the inference result corresponding to at least one network element (i.e., including the inference result of the first model). For example, the first inference result may be an aggregated result of the inference result corresponding to at least one network element. Specifically, for example, the first inference result may be a weighted sum of the inference results corresponding to at least one network element, or may be the product of the inference results corresponding to at least one network element. The embodiments of the present application do not limit the manner in which the first inference result is determined.

[0314] The following introduces the content of determining the first reasoning result by taking as an example at least one network element including a first network element and a third network element, the first network element corresponding to the first model, and the third network element corresponding to the third model.

[0315] Exemplarily, the second network element may aggregate the reasoning result of the first model and the reasoning result of the third model to obtain the first reasoning result.

[0316] In the case that the inference result of the third model includes multiple inference results, the second network element may respectively determine the first inference result corresponding to each of the multiple inference results of the third model, that is, obtain multiple first inference results.

[0317] For example, the second network element may calculate a first inference result based on one inference result from the multiple inference results of the third model and an inference result from the multiple inference results of the first model that has the same order as the one inference result, and so on, thereby obtaining multiple first inference results.

[0318] For example, at least one network element includes a first network element and a third network element. The second information fed back by the first network element indicates reasoning result 1, reasoning result 2, and reasoning result 3, and the second information indicates that the order of the indicated reasoning results is: reasoning result 1, reasoning result 2, and reasoning result 3. The second information fed back by the third network element indicates reasoning result 4, reasoning result 5, and reasoning result 6, and the order of the indicated reasoning results is: reasoning result 6, reasoning result 5, and reasoning result 4. Then, the second network element can determine a first reasoning result based on reasoning result 1 of the first network element and reasoning result 4 of the second network element, determine a first reasoning result based on reasoning result 2 and reasoning result 5, and determine a first reasoning result based on reasoning result 3 and reasoning result 6. That is, the first reasoning result includes the three determined first reasoning results.

[0319] With the second network element as a participant in distributed reasoning, the first information includes first indication information or information about the first condition, at least one network element includes only one network element (i.e., the first network element), and taking the first network element determining the first data with reference to the above-mentioned method C1 as an example, in conjunction with the schematic diagram of a communication method shown in FIG8 , an example of the interaction between the various network elements involved in FIG7 is introduced. The various steps involved in FIG8 are introduced below.

[0320] S801. The second network element receives a service prediction request.

[0321] A service prediction request, also known as a service processing request, is used to indicate the processing or prediction of a service result corresponding to a service. For example, the service prediction request may carry an analysis identifier, which is used to identify the requested analysis type. For example, an analysis identifier of Service Experience requests statistical or predicted results of the service experience. The second network element receives the service prediction request, for example, from a terminal device or other network element.

[0322] In another possible implementation, the second network element may have previously received a service prediction request, or the second network element may have independently determined that it needs to process a certain service. In these cases, the second network element does not need to receive the service prediction request. That is, S801 is an optional step, which is indicated by a dotted line in FIG9 .

[0323] S802: The second network element sends first information to the first network element. Correspondingly, the first network element receives the first information from the second network element.

[0324] In one possible implementation, the first information includes first indication information. The first indication information is used to determine the time of inference data. In another possible implementation, the first information includes first condition information. The first condition information is used to determine the time of inference data.

[0325] The content of the first information, the content of the first indication information, the content of the first condition information, and the content of the first indication information used to determine the time of inference data can be respectively referred to the content of the first information, the content of the first indication information, the content of the first condition information, and the content of the first indication information used to determine the time of inference data discussed in Figure 7 above, and the repetitions are not listed here again.

[0326] S803. The first network element determines an inference result of the first model based on the first data.

[0327] The first network element may determine the first data based on the time used to determine the inference data, or may be described as determining the first data based on the first information. The content of determining the first data can refer to the content of determining the first data discussed in FIG. 7 above, and will not be listed here. After determining the first data, the first network element may determine the inference result of the first model based on the first data. The content of determining the inference result of the first model can refer to the content of determining the inference result of the first model discussed in FIG. 7 above, and will not be repeated here.

[0328] S804: The first network element sends second information to the second network element. Correspondingly, the second network element receives the second information from the first network element. The second information indicates the inference result of the first model.

[0329] The content of the second information and the way in which the second information indicates the reasoning result of the first model can refer to the content of the second information and the content of the second information indicating the reasoning result of the first model discussed in Figure 8 above, and will not be listed here.

[0330] S805. The second network element determines an inference result of the second model of the second network element based on the second data.

[0331] The second network element may determine the second data based on the time used to determine the inference data. The method for determining the second data can refer to the determination of the second data content discussed in FIG. 8 above and will not be further detailed here. The content of the inference result of the second model and the method for determining the content of the inference result of the second model can refer to the content of the inference result of the second model and the method for determining the content of the inference result of the second model discussed in FIG. 7 above and will not be further detailed here.

[0332] If the second network element is the coordinator of the distributed reasoning, the second network element may not need to perform step S805 , that is, step S805 is an optional step, which is indicated by a dotted line in FIG8 .

[0333] S806. The second network element determines a first reasoning result based on the reasoning result of the first model and the reasoning result of the second model.

[0334] The content of the first reasoning result and the method for determining the first reasoning result can refer to the content of the first reasoning result and the method for determining the first reasoning result discussed in FIG. 7 above, and will not be listed here.

[0335] When the second network element is the coordinator of distributed reasoning, the at least one network element can be multiple network elements, each of which can execute the process performed by the first network element. Accordingly, the second network element can also determine the first reasoning result based on the reasoning results from the multiple network elements. In this case, the interaction process between the second network element and the multiple network elements involved can refer to the interaction process between the second network element and the first network element discussed in Figure 9 above, and will not be listed here one by one.

[0336] Taking the second network element as a participant in distributed reasoning and the first network element determining the first data with reference to the above-mentioned C2 method as an example, the interaction between the various network elements involved in Figure 7 is described by way of example in conjunction with the schematic diagram of a communication method shown in Figure 9. The following describes the various steps illustrated in Figure 9.

[0337] S901. The second network element receives a service prediction request.

[0338] The content of the service prediction request may refer to the content of the service prediction request discussed above in Figure 8. S901 is an optional step, which is indicated by a dotted line in Figure 9.

[0339] S902: The second network element sends first information to the first network element. Correspondingly, the first network element receives the first information from the second network element. The first information is used to determine the time of inference data.

[0340] The content of the first information can refer to the content of the first information discussed in Figure 7 above, and the content of the first information used to determine the time of inference data can also refer to the content of the first information used to determine the time of inference data discussed in Figure 7 above, which will not be listed here.

[0341] S903. The first network element determines at least one data based on the first information.

[0342] The first network element may determine the content of at least one data according to the first information with reference to the discussion of determining the content of at least one data according to the first information in FIG. 7 , which will not be enumerated here.

[0343] S904: The first network element sends information about at least one data item to the second network element. Correspondingly, the second network element receives information about at least one data item from the first network element. The information about at least one data item can refer to the information about at least one data item discussed in FIG. 7 above and is not further detailed here.

[0344] S905: The second network element sends fourth information to the first network element. Correspondingly, the first network element receives the fourth information from the second network element. The fourth information indicates that the first data should be used for model inference.

[0345] The second network element can determine the first data from the at least one data. The content of the first data can refer to the second network element determining the content of the first data discussed in FIG8 above, which is not listed here. The content of the fourth information can refer to the content of the fourth information discussed in FIG7 above.

[0346] S903 to S905 are optional steps, which are indicated by dotted lines in FIG9 .

[0347] S906. The first network element determines an inference result of the first model of the first network element based on the first data.

[0348] The content of determining the inference result of the first model can refer to the content of determining the inference result of the first model discussed in FIG. 8 , and will not be repeated here.

[0349] S907: The first network element sends second information to the second network element. Correspondingly, the second network element receives the second information from the first network element. The second information indicates the inference result of the first model.

[0350] The content of the second information and the way in which the second information indicates the reasoning result of the first model can refer to the content of the second information and the content of the second information indicating the reasoning result of the first model discussed in Figure 8 above, and will not be listed here.

[0351] S908. The second network element determines an inference result of the second model of the second network element based on the second data.

[0352] The content of the second data, determining the content of the second data, and determining the content of the reasoning result of the second model can refer to the content of the second data, determining the content of the second data, and determining the content of the reasoning result of the second model discussed in Figure 8 above, and will not be listed here.

[0353] S908 is an optional step, indicated by a dotted line in FIG9 .

[0354] S909. The second network element determines a first reasoning result based on the reasoning result of the first model and the reasoning result of the second model.

[0355] The second network element determines the content of the first reasoning result, and the content of the first reasoning result can refer to the determination of the content of the first reasoning result and the content of the first reasoning result discussed in Figure 7 above, respectively, and are not listed here.

[0356] In an embodiment of the present application, the second network element can determine the first data from at least one data determined based on the first information. In this way, the second network element can combine the information of the data of each network element participating in the model reasoning to determine more reasonable reasoning data, thereby ensuring the accuracy of the model reasoning.

[0357] In the process of training the model, if each participant does not consider the time of training data, it will also affect the accuracy of model training.

[0358] In view of this, an embodiment of the present application provides a communication method, in which a coordinator or initiator (such as a second network element) can indicate to other participants (such as a first network element) the time of training data used to determine model training, so that each participant can use training data that meets the time to perform model training to ensure the accuracy of model training.

[0359] Please refer to Figure 10 for a communication method provided in an embodiment of the present application. This embodiment of the present application takes multiple network elements participating in model training as an example, where these multiple network elements include a first network element. Optionally, the multiple network elements may also include a second network element and / or a third network element, and the number of these multiple network elements is not specifically limited. The following describes the various steps illustrated in Figure 10.

[0360] S1001: A second network element sends first information. This embodiment of the present application takes at least one network element receiving first information from a second network element as an example. The first information indicates a time for determining training data. The at least one network element does not include the second network element.

[0361] The at least one network element belongs to some or all of the multiple network elements. When the second network element is a coordinator of distributed learning, the at least one network element may be all of the multiple network elements. When the second network element is a participant in distributed learning, the at least one network element may be a network element in the multiple network elements other than the first network element.

[0362] In one possible implementation, the second network element may broadcast first information, where the first information carries information about distributed training, where the information about distributed training includes, for example, a distributed training ID and / or IDs of multiple models. The distributed training ID and the IDs of multiple models may be the same as those discussed above, and are not further detailed here.

[0363] Multiple models are associated with a distributed learning ID. Multiple models can be regarded as models participating in distributed learning, and multiple models correspond to multiple network elements. For example, multiple models correspond one-to-one with multiple network elements. The content of the one-to-one correspondence between multiple models and multiple network elements can refer to the one-to-one correspondence between multiple models and multiple network elements discussed in Figure 7 above, and will not be listed here. In the embodiment of the present application, the models other than the model of the first network element (such as the first model) in the multiple models are referred to as one or more models. The one or more models may include a model of the second network element, and / or a model of the third network element, etc.

[0364] In another possible implementation, the second network element may send first information to at least one network element, where the first information notifies the at least one network element of the time for training data. In this implementation, the first information sent to one of the at least one network elements (e.g., the first network element) may optionally include the ID and / or distributed learning ID of the model of the first network element. The ID of the model of the first network element is one of the multiple model IDs discussed above.

[0365] The relationship between at least one network element and multiple models can refer to the relationship between at least one network element and multiple models discussed in Figure 7 above, and will not be listed here.

[0366] The following describes the content of the first information using an example in which at least one network element includes a first network element, and a second network element sends first information to the first network element. The first information includes at least one item of information from H1 to H5, indicating the time used to determine the inference data. Each of these items is described below. It is understood that at least one item from H1 to H5 can be considered an implementation method for indicating the time used to determine the inference data.

[0367] H1. The first information includes information of a first timestamp. The information of the first timestamp is used to determine the time of the training data.

[0368] The information of the first timestamp can refer to the content of the first timestamp information discussed in Figure 7 above and will not be listed here. Unlike the information content of the first timestamp information discussed in Figure 7 above, the information of the first timestamp involved in the embodiment of the present application is the information of the timestamp corresponding to the training data. For example, if the content of the first timestamp information discussed in Figure 7 is replaced with "training", the content of the first timestamp information in the embodiment of the present application can be obtained.

[0369] For example, the information of the first timestamp includes information of the first moment and / or time offset. The first moment includes a related moment of collecting (or called collecting) training data. Time offset refers to the deviation of the time allowing the collection or generation of training data, or can be described as a time deviation indicating the actual collection or generation time of acceptable training data relative to the first timestamp. The content of the information of the first moment and the time offset can also refer to the content of the information of the first moment and the time offset discussed in Figure 7 above, respectively, and are not listed here. Different from the content of the first moment and the time offset discussed in Figure 7 above, the first moment and the time offset involved in the embodiment of the present application are the first moment and the time offset corresponding to the training data. For example, if the "inference" in the content of the information of the first moment and the time offset discussed in Figure 7 is replaced with "training", the information of the first moment and the time offset in the embodiment of the present application can be obtained.

[0370] H2. The first information includes information of the first time window.

[0371] The content of the first time window can be referred to as the content of the first time window discussed in FIG. 7 above, and the content of the first time window, respectively, and will not be listed here. Unlike the content of the first time window discussed in FIG. 7 above, the content of the first time window involved in the embodiment of the present application is the time window corresponding to the training data. For example, if the word "inference" in the content of the first time window discussed in FIG. 7 is replaced with "training", the content of the first time window in the embodiment of the present application can be obtained.

[0372] In one possible design, the first information includes information about the first timestamp and information about the first time window. The information about the first timestamp can refer to the content of the first timestamp information discussed in H1 above, and the information about the first time window can refer to the content of the first time window information discussed in H2 above. Any repetitions are not listed here. For example, the first information can include the start time, first period, and time offset of the first time window. Alternatively, the first information can include the end time, first period, and time offset of the first time window.

[0373] H3. The first information includes information about the time when model-based training starts.

[0374] The content of the moment of starting model-based training can refer to the content of the moment of starting model-based reasoning discussed in Figure 7 above, and will not be listed here. Unlike the moment of starting model-based reasoning discussed in Figure 7 above, the embodiment of the present application involves the moment of starting model-based training. For example, if "reasoning" in the content of the moment of starting model-based reasoning discussed in Figure 7 is replaced with "training", the content of the moment of starting model-based training in the embodiment of the present application can be obtained.

[0375] The first network element can determine the time for determining the training data based on the moment when the model-based training is started. The content for determining the time for determining the training data based on the moment when the model-based training is started can also refer to the content for determining the time for inference data based on the moment when the model-based reasoning is started as discussed in FIG7 above, which will not be listed here. For example, if "reasoning" is replaced with "training" in the content for determining the time for inference data based on the moment when the model-based reasoning is started as discussed in FIG7, the content for determining the time for inference data based on the moment when the model-based reasoning is started in the embodiment of the present application can be obtained.

[0376] H4. The first information includes information about the first condition.

[0377] The first condition may also be referred to as a training trigger condition, which indicates the conditions under which the first network element triggers the execution of training, that is, under which conditions (such as certain events) the first network element triggers the first model training.

[0378] Optionally, the content of the first condition can refer to the content of the first condition discussed in FIG. 7 above and will not be listed here.

[0379] H5. The first information includes first indication information.

[0380] The first indication information indicates that training is to be performed based on the model, and the first indication information may also be referred to as a model training indication. The content of the first indication information can refer to the content of the first indication information discussed in FIG. 7 above, and will not be listed here. Unlike the first indication information discussed in FIG. 7 above, the embodiment of the present application involves the first indication information being the first indication information corresponding to the training data. For example, if "inference" in the content of the first indication information discussed in FIG. 7 is replaced with "training", the content of the first indication information in the embodiment of the present application can be obtained.

[0381] In another possible implementation, when the first information is the information shown in H1 or H2 above, the first information may further include at least one of information about the time when model-based training is started, information about the first condition, or first indication information. In this case, at least one of the information about the time when model-based training is started, information about the first condition, or first indication information is not used to determine the time of the training data.

[0382] Optionally, the first information may also indicate information about a mode for sending training results. For example, the mode for sending training results indicates whether the training results are sent periodically or aperiodically. In this way, at least one network element may determine the timing of sending the training results based on the mode for sending training results indicated by the first information. The training result may be the result of encrypting the updated local parameters and / or loss function according to an encryption algorithm, etc. The encryption algorithm may be determined by negotiation among the participating parties, or may be pre-configured in at least one network element, and there is no limitation on this. The updated local parameters refer to the local parameters obtained after updating the model based on the training data, and the loss function refers to the loss function obtained by solving this training.

[0383] As for the content of the first information sent by the second network element to at least one network element other than the first network element, reference may be made to the content of the above-mentioned first information, and will not be listed one by one here.

[0384] S1002. The first network element determines a training result of the first model based on the first data and the first model.

[0385] After receiving the first information, the at least one network element may determine training data according to the first information. Taking the example that the at least one network element includes a first network element and the model of the first network element is a first model, the content of S1002 is introduced by way of example.

[0386] After receiving the first information, the first network element can determine the first data based on the first information. The method for determining the first data can refer to the determination of the content of the first data discussed in Figure 7 above, and will not be listed here. The first network element can determine the input of the first model based on the first data. Optionally, the first data is the input of the first model, or the first data is preprocessed to obtain the input of the first model. The content of the preprocessing can refer to the content of the preprocessed data discussed above, and will not be listed here. The input of the first model can be in the form of a vector, an image, a matrix, etc., which is not specifically limited.

[0387] The first network element inputs the input of the first model into the first model to obtain the input of the first model, and determines the training result of the first model based on the output of the first model. The training result of the first model includes the loss function corresponding to the first model and / or the updated local parameters, specifically, for example, the result of encrypting the loss function corresponding to the first model according to the first encryption algorithm, and / or the result of encrypting the updated local parameters. For example, the first network element can determine the loss function of the first model based on the output of the first model and the label corresponding to the first data. The first network element can also adjust the parameters of the first model based on the loss function of the first model to obtain the updated local parameters of the first model.

[0388] For example, the first network element negotiates with the coordinator to perform a first encryption algorithm, or the coordinator determines the first encryption algorithm. For example, the first encryption algorithm is a homomorphic encryption algorithm, and the coordinator can determine the encryption key pair corresponding to the homomorphic encryption algorithm, and the encryption key pair includes a public key and a private key. The coordinator can send the public key to each network element participating in the model training (such as the first network element). The first network element can encrypt the loss function of the first model and / or the updated parameters of the first model based on the public key to obtain the training result of the first model.

[0389] If the second network element is a coordinator, the second network element may specify the first encryption algorithm. If the second network element is a participant, the coordinator may also send the public key to the second network element. In this way, the second network element may also obtain the public key.

[0390] For example, the initial value of the local parameter of the first model is Θ A , then the first network element determines the updated local parameters of the first model as The loss function of the first model is L A The first network element encrypts the updated local parameters of the first model according to the public key to obtain the encrypted local parameters Furthermore, the first network element can encrypt the loss function of the first model according to the public key to obtain the encrypted loss function

[0391] Optionally, when the first data includes multiple data, the first network element may obtain multiple training results based on the multiple data and the first model. In other words, the training result of the first model may include multiple training results. Each of the multiple training results may include an updated local parameter and / or loss function.

[0392] The following describes an example of how the first network element determines the first data in conjunction with the implementation shown in K1 or K2.

[0393] K1. The data determined by the first network element according to the first information includes at least one data, and the at least one data is used as the first data.

[0394] K2. The first network element determines, based on the notification from the second network element, to use the first data among the at least one data as training data.

[0395] Exemplarily, the data determined by the first network element based on the first information includes at least one data, and the first network element sends information about the at least one data to the second network element. The second network element can determine the first data from the at least one data and send fourth information to the first network element, where the fourth information instructs the first network element to use the first data for training, or can be described as instructing the first network element to use the first data as training data. Under this embodiment, since the second network element combines information about the data fed back by at least one network element to determine the training data used by at least one network element to participate in model training, this is conducive to more accurate determination of the training data.

[0396] The information of at least one data item is used to indicate the at least one data item. The information of at least one data item includes information about each data item in part or all of the at least one data item. The at least one data item includes first data. The information of each data item is described below using the information of the first data item as an example. The information of the first data item includes at least one item of information from L1 to L5. The information shown in L1 to L5 is described below.

[0397] L1. Information on the serial number of the first data.

[0398] The content of the first data number and the method of determining the first data number can refer to the content of the first data number and the method of determining the first data number discussed in FIG. 7 , respectively.

[0399] Under L1, the content of the method for determining the first data and the content of the fourth information by the first network element can refer to the content of the method for determining the first data and the content of the fourth information discussed in D1 of Figure 7 above, and are not listed here.

[0400] For example, the second network element may use the smallest-numbered data among the at least one data corresponding to each model in the models participating in the distributed training as the training data corresponding to each network element. In this case, the first data may be the smallest-numbered data among the at least one data corresponding to the first network element. Alternatively, the second network element may use the largest-numbered data among the at least one data corresponding to each network element in at least one network element as the training data corresponding to each network element. In this case, the first data may be the largest-numbered data among the at least one data corresponding to the first network element.

[0401] Alternatively, the second network element may determine as the first data data the data whose number difference between the at least one data corresponding to the first network element and the training data of one or more models is less than or equal to the first threshold. In this case, the difference between the number of the first data and the number of the training data of one or more models is less than or equal to the first threshold. The method for the second network element to obtain the first threshold can refer to the content of the second network element obtaining the first threshold discussed in Figure 7 above, and will not be further detailed here.

[0402] L2. Data characteristic information: The data characteristic information indicates the characteristics (or attributes) of the first data.

[0403] The content of the data feature information can refer to the data feature information discussed in FIG. 7 above and will not be repeated here. Feature information, for example, includes a feature identifier, which is used to indicate the feature. The content of the feature identifier can refer to the content of the feature identifier discussed in FIG. 7 above and will not be repeated here.

[0404] Under L2, the content of the method for determining the first data and the content of the fourth information by the first network element can refer to the content of the method for determining the first data and the content of the fourth information discussed in D2 of Figure 7 above, and are not listed here.

[0405] Optionally, the fourth information may include data feature information corresponding to the first data and / or information on the serial number of the first data, and the data feature information and / or information on the serial number of the first data are used to instruct the first network element to use the first data as training data.

[0406] L3. First time information: The first time information indicates the time when the first data is actually collected or generated.

[0407] The first time information includes, for example, information about a second timestamp and / or information about a second time window. The second timestamp information indicates the actual time when the training data was collected or generated, and includes, for example, a second time. The second time is, for example, the actual time when the first data was collected, the actual time when the first data was generated, the actual time when the first data was generated, or the actual time when the first data was generated.

[0408] The second time window is the time window for actually collecting or generating the first data. The content of the second time window can refer to the content of the first time window discussed in FIG10 above, and will not be listed here.

[0409] Under L3, the content of the method for determining the first data and the content of the fourth information by the first network element can refer to the content of the method for determining the first data and the content of the fourth information discussed in D3 of Figure 7 above, and are not listed here.

[0410] In one possible implementation, the second network element may determine as first data data the time difference between at least one data corresponding to the first network element and the training data corresponding to one or more models being less than or equal to a second threshold. In this case, the time difference between the actual acquisition or generation time of the first data and the training data corresponding to the one or more models being less than or equal to the second threshold. The second network element may also obtain the content of the second threshold by referring to the discussion of the second network element obtaining the content of the first threshold in FIG10 above, and will not be further detailed here.

[0411] In the case where the first time information corresponding to the first network element indicates a second time window, the time difference between the first data and the training data corresponding to one or more models may be the difference between the start times of the two second time windows corresponding to the first data and the training data corresponding to the one or more models, or the difference between the end times of the two second time windows, or the sum of the difference between the start times of the two second time windows and the difference between the end times of the two second time windows, etc., and no specific limitation is given to this.

[0412] In another possible implementation, the second network element may determine as the first data the data for which the sum of the time of the training data corresponding to the one or more models in the at least one data corresponding to the first network element is less than or equal to the minimum. In this case, the sum of the time of the actual collection or generation of the first data and the time of the training data corresponding to the one or more models is the minimum. The meaning of the one or more models can refer to the content of the one or more models involved in D1 above and will not be listed here.

[0413] Optionally, the fourth information may include second time information corresponding to the first data and / or information on the number of the first data, where the second time information and / or information on the number of the first data are used to instruct the first network element to use the first data as training data.

[0414] L4. First location information: The first location information indicates the location where the first data is collected.

[0415] For example, if the first data is collected in a certain area (such as an area of ​​interest), the location of the first data is the area. Alternatively, if the first data is collected in a certain location, the location is the location of the first data.

[0416] Under L4, the content of the method for determining the first data and the content of the fourth information by the first network element can refer to the content of the method for determining the first data and the content of the fourth information discussed in D4 of Figure 7 above, and are not listed here.

[0417] In one possible implementation, the second network element may determine as first data data whose distance between the location of the at least one data corresponding to the first network element and the location of the training data corresponding to one or more models is less than or equal to a third threshold. In this case, the distance between the location of the first data and the location of the training data corresponding to the one or more models is less than or equal to the third threshold. The content of the second network element obtaining the third threshold can refer to the content of the second network element obtaining the first threshold described above and is not further detailed here.

[0418] L5. Access type information. The access type information indicates the access type corresponding to the first data. For example, the access type corresponding to the first data is 4G or 5G.

[0419] Under L5, the content of the method for determining the first data and the content of the fourth information by the first network element can refer to the content of the method for determining the first data and the content of the fourth information discussed in D5 of Figure 7 above, and are not listed here.

[0420] For example, the second network element may determine, as the first data, data of the same access type as the training data corresponding to the one or more models in the at least one data corresponding to the first network element. In this case, the location of the first data is the same as the access type corresponding to the training data corresponding to the one or more models.

[0421] If the information of the at least one data corresponding to the first network element includes at least two items of information from L1 to L5, the second network element may also determine the first data from the at least one data corresponding to the first network element based on one of the at least two items of information. Alternatively, the second network element may determine the first data from the at least one data corresponding to the first network element based on the contents of the at least two items of information.

[0422] For example, if the information of at least one data corresponding to the first network element includes the above-mentioned L1 (i.e., the data number), L3 (i.e., the first time information) and L5 (i.e., the access type information), then the second network element can determine the data in the at least one data corresponding to the first network element, whose time difference with the training data corresponding to one or more models is less than or equal to the second threshold, as the first data.

[0423] For another example, if the information of at least one data corresponding to the first network element includes the above-mentioned L3 (i.e., the first time information), L4 (i.e., the first location information), and L5 (i.e., the access type information), then the second network element can determine the at least one data corresponding to the first network element, the data whose time difference with the training data corresponding to one or more models is less than or equal to the first threshold, the difference in location distance is less than or equal to the third threshold, and the access type is the same as the first data. Alternatively, the second network element can also determine a piece of data for the model for client B and client C based on other principles (such as location and access type, etc.). Other principles include that the initiator determines a piece of data for client B and client C respectively, and the location and access type are the same, and the data are collected as close as possible in time.

[0424] In another possible implementation, the second network element determines that there is no first data meeting the corresponding condition in at least one data corresponding to the first network element, or it can be described as that the first network element does not have data that meets the time for model training, then the first network element may send a second indication message to the at least one network element. The second indication message indicates the termination of model training. Optionally, the second indication message may carry a cause value for model training termination (model training termination with cause code), which, for example, indicates that some participants do not have data that meets the requirements of distributed training.

[0425] The role of the second network element in the distributed training process is different, and the subsequent processing process is also different, which is introduced below.

[0426] In one possible case, when the second network element is a participant in distributed training, steps S1003 to S1004 may be performed.

[0427] S1003: The first network element sends second information to the second network element. Accordingly, the second network element receives the second information from the first network element. The second information indicates the training result of the first model. Alternatively, the second information may include information about the training result of the first model.

[0428] After obtaining the second information, the first network element may send the second information to the second network element. The first network element may send the second information to the second network element directly, or the first network element may send the second information to the second network element through coordination, which is not specifically limited.

[0429] Similarly, the second network element can obtain the training results of the model fed back by at least one network element to obtain the training results corresponding to at least one network element. Optionally, if the second network element is a participant in the distributed training, the second network element can also determine the second data based on the first information.

[0430] S1004. The second network element determines the training result of the second model based on the second data and the second model. The second data can be determined by referring to the determination of the first data, and the training result of the second model can be determined by referring to the determination of the first model, which are not listed here.

[0431] Continuing with the example of the first network element and the second network element negotiating that the first encryption algorithm is a homomorphic encryption algorithm, the initial value of the local parameter of the second model is Θ B The second network element determines that the updated local parameters of the second model are The loss function of the second model is L B The second network element uses the public key to encrypt the updated local parameters of the second model to obtain the encrypted result of the updated local parameters. And obtain the encrypted result of the loss function of the second model

[0432] In this way, the second network element can determine the first training result based on the training results corresponding to at least one network element (i.e., including the training results of the first model) and the training results of the second model. For example, the first training result can be an aggregated result of the training results corresponding to at least one network element and the training results of the second model. Specifically, for example, the first training result can be a weighted sum of the training results corresponding to at least one network element and the training results of the second model, or can be the product of the training results corresponding to at least one network element and the training results of the second model. The embodiment of the present application does not specifically limit the content of determining the first training result.

[0433] The following takes at least one network element including a first network element as an example to introduce the content of determining the first training result.

[0434] Exemplarily, the second network element may aggregate the training results of the first model and the training results of the second model to obtain a first training result. The first training result includes the loss functions of multiple models and / or the updated local parameters of multiple models, specifically, for example, including the encrypted results of the loss functions of multiple models and / or the encrypted results of the updated local parameters of multiple models. The loss functions of multiple models may be the sum of the loss functions of multiple models. The updated local parameters of multiple models may be, for example, the sum of the updated local parameters of multiple models.

[0435] For example, continuing to take the first network element and the second network element negotiating that the first encryption algorithm is a homomorphic encryption algorithm as an example, the training result of the first model includes the encryption result of the local parameters after the first model is updated The encrypted result of the loss function of the first model is The encrypted result of the updated local parameters of the second model is And the encrypted result of the loss function of the second model is In this way, the second network element can be based on the local parameters of the slave party. and loss function And the loss function etc., determine the loss functions of multiple models And the encryption results of the updated local parameters of the multiple models can also be determined based on the updated local parameters of the second network element and the updated local parameters of the first network element Intermediate parameters For example it could be y i The labels for model training are the actual target results corresponding to the model input data, such as the aggregated results of the labels of multiple models.

[0436] If the training results of the second model include multiple training results, the second network element may determine the first training result corresponding to each of the multiple training results of the second model, i.e., obtain multiple first training results. For example, the second network element may calculate a first training result based on one of the multiple training results of the second model and training results of the multiple training results of the first model that are in the same order as the one training result. This process can be repeated and so on, thereby obtaining multiple first training results.

[0437] In addition to the information of the training result of the first model, the second information also includes the first time information and / or the order of multiple training results included in the training result of the first model, etc. The content of the first time information can refer to the content of the first time information discussed above and is not listed here.

[0438] When the second information also includes the first time information, the second network element may optionally verify whether the training results of the first model are appropriate based on the first time information. Verification of the appropriateness of the training results of the first model can refer to the verification of the appropriateness of the inference results of the first model discussed in FIG. 7 above and will not be repeated here. Because the training results of the first model are further verified based on the first time information, the accuracy of model training can be further improved.

[0439] When the training results of the first model include multiple training results, these multiple training results correspond to multiple data in the first data. For example, one training result corresponds to one data in the multiple data in the first data, or it can be described as a training result determined based on one data in the multiple data in the first data and the first model. The order and acquisition method of these multiple training results can refer to the order and acquisition method of the multiple training results discussed in Figure 7 above, and will not be enumerated here.

[0440] In one possible case, when the second network element is a coordinator of distributed training, step S1005 may be executed.

[0441] S1005. The second network element receives the fifth information. For example, the second network element receives the fifth information from the coordinator of the distributed training. The fifth information indicates the first training result. The second network element can determine whether to end the distributed training process based on the first training result, that is, determine whether to end the training of multiple models or continue the training of multiple models. If the second network element determines that the first training result meets the conditions, the distributed training process can be ended; if the second network element determines that the first training result does not meet the conditions, the distributed training process can be continued. The conditions here can be, for example, the convergence of the loss function corresponding to the first training result, or the number of training times reaches a preset number, etc., which are not limited.

[0442] If the distributed training process is stopped, the second network element may send a second indication message to at least one network element respectively, and the second indication message indicates the end of the distributed training. Let's take the example where the first network element and the second network element adopt the homomorphic encryption algorithm, and the second network element is the coordinator. The second network element uses the private key to decrypt the first training result to obtain the decrypted first training result, and then determines whether to stop training the multiple models based on the first training result. Alternatively, if the distributed training process is not ended, the second network element may send a third indication message to at least one network element respectively, and the third indication message sent to one of the at least one first network element may also indicate the global gradient of the local parameters of the one network element. And so on, until the distributed training is ended.

[0443] Optionally, after obtaining the first training result, the second network element may further send the first training result to at least one network element. This allows the at least one network element to determine the encrypted global gradient of the local parameter based on the first training result and the local parameter corresponding to each network element. The global gradient of the local parameter of a network element refers to the updated value of the local parameter of the network element. The at least one network element may further send the encrypted global gradient of the local parameter to the second network element. The second network element decrypts the encrypted global gradient of the local parameter corresponding to each of the at least one network element, thereby obtaining the global gradient of the local parameter corresponding to each of the at least one network element.

[0444] For example, the first network element calculates and updates the gradient of the local parameter And add random noise Get the encrypted global gradient, that is The encrypted gradient is sent to the second network element. The second network element uses the private key to decrypt the encrypted local gradient and obtain the decrypted global gradient of the local parameter, that is, The second network element sends the global gradient of the local parameter to the first network element. The first network element subtracts random noise from the global gradient of the local parameter to obtain the decrypted global gradient of the local parameter. And update the local parameters based on the global gradient of the decrypted local parameters to obtain the updated local parameters Θ A Similarly, the initiator (such as the third network element) calculates the gradient of the updated local parameters Get the encrypted global gradient, that is The encrypted gradient is sent to the second network element. The second network element uses the private key to decrypt the encrypted local gradient and obtain the decrypted global gradient of the local parameter, that is, The second network element sends the global gradient of the local parameter to the first network element. The first network element subtracts random noise from the global gradient of the local parameter to obtain the decrypted global gradient of the local parameter. And update the local parameters based on the global gradient of the decrypted local parameters to obtain the updated local parameters Θ B .

[0445] S1003 - S1004 and S1005 are two optional parts, that is, they are optional steps, and are indicated by dotted lines in FIG10 .

[0446] In the embodiments of the present application, each network element participating in model training can, based on the instructions of a second network element, clearly determine the time used for training data. This allows each network element to perform model training based on training data with similar or identical time periods, which is beneficial for improving the accuracy of model training. Furthermore, since the participants in distributed learning exchange intermediate computational results during training, no participant can access the other's original data throughout the training process. Instead, the model training process is completed by exchanging intermediate results. This not only achieves joint training but also protects the privacy of the participants' local data.

[0447] When there are three or more participants, the distributed training process involved can be referenced to the distributed learning process described in FIG10 , and will not be listed here. Distributed learning can involve a variety of model training methods, such as multi-party multi-classification distributed learning based on privacy-preserving label sharing and multi-party distributed learning based on secret sharing, and other algorithms, without specific limitation.

[0448] It is understood that, in order to implement the functions in the above embodiments, the base station and the terminal include hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily appreciate that, in conjunction with the units and method steps of the various examples described in the embodiments disclosed in this application, this application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in hardware or in a computer software-driven hardware manner depends on the specific application scenario and design constraints of the technical solution.

[0449] Please refer to Figure 11, which is a schematic diagram of the structure of a communication device provided in an embodiment of the present application. The communication device can be used to implement the functions of the first network element or the second network element in the above method embodiment, and thus can also achieve the beneficial effects possessed by the above method embodiment. In the embodiment of the present application, the communication device can be the master participant, coordinator or slave participant involved in Figure 2, the master participant, coordinator or slave participant involved in Figure 3, the coordinator, network element D, other participants, network element B or network element C involved in Figure 4, the initiator or network element A involved in Figure 4, or any network element or VFL Server involved in Figure 5, or any network element or VFL Server involved in Figure 6, etc., and can also be a module (such as a chip) in these network elements.

[0450] As shown in Figure 11, the communication device 1100 includes a processing module 1110 and a transceiver module 1120. The communication device 1100 is used to implement the functions of the first network element or the second network element in any of the method embodiments shown in Figures 7 to 10 above.

[0451] As an embodiment, the communication device 1100 is used to implement the function of the second network element in any one of the method embodiments shown in Figures 7 to 9.

[0452] Illustratively, the transceiver module 1120 is configured to transmit the first information and receive the second information, and the processing module 1110 is configured to determine the first reasoning result, etc. The contents of the first information, the second information, and the second reasoning result can be referred to as the first information, the second information, and the second reasoning result involved in any of the method embodiments shown in FIG. 7 to FIG. 9 , respectively, and are not listed one by one here.

[0453] The communication device 1100 can also implement the steps performed by the second network element in any of the method embodiments of Figures 7 to 9 above, which are not listed one by one here.

[0454] As an embodiment, the communication device 1100 is used to implement the function of the first network element in any one of the method embodiments shown in Figures 7 to 9.

[0455] Exemplarily, the transceiver module 1120 is configured to receive the first information and send the second information, and the processing module 1110 is configured to determine the inference result of the first model. The first information, the second information, and the content of determining the inference result of the first model can be referred to as the first information, the second information, and the content of determining the inference result of the first model involved in any of the method embodiments shown in Figures 7 to 9 above, respectively, and will not be enumerated one by one here.

[0456] The communication device 1100 can also implement the steps performed by the first network element in any of the method embodiments of Figures 7 to 9 above, which are not listed one by one here.

[0457] As an embodiment, the communication device 1100 is used to implement the function of the second network element in the method embodiment involved in Figure 10.

[0458] Exemplarily, the first information is sent, etc. The content of the first information can refer to the content of the first information involved in the method embodiment involved in FIG10 above, and will not be listed one by one here.

[0459] The communication device 1100 can also implement the steps performed by the second network element in the method embodiment involved in Figure 10 above, which are not listed one by one here.

[0460] As an embodiment, the communication device 1100 is used to implement the function of the first network element in the method embodiment involved in Figure 10.

[0461] Exemplarily, first information is received, and the training result of the first model is determined, etc. The first information and the content of determining the training result of the first model can be respectively referred to the first information involved in the method embodiment involved in FIG. 10 and the content of determining the training result of the first model, and are not listed one by one here.

[0462] The communication device 1100 can also implement the steps performed by the first network element in the method embodiment involved in Figure 10 above, which are not listed one by one here.

[0463] A more detailed description of the processing module 1110 and the transceiver module 1120 can be directly obtained by referring to the relevant descriptions in the method embodiments shown in Figures 7 to 10, and will not be repeated here.

[0464] Please refer to Figure 12, which is a structural diagram of a communication device provided in an embodiment of the present application. As shown in Figure 12, the communication device 1200 includes a processor 1210 and an interface circuit 1220. The processor 1210 and the interface circuit 1220 are coupled to each other. It will be understood that the interface circuit 1220 can be a transceiver or an input / output (I / O) interface. Optionally, the communication device 1200 may further include a memory 1230 for storing instructions executed by the processor 1210 or storing input data required for the processor 1210 to run instructions or storing data generated after the processor 1210 runs instructions. Optionally, the processor 1210 can deploy one or more models, such as the first model or the second model mentioned above. For example, the processor includes an AI chip, and the one or more models can be deployed on the AI ​​chip.

[0465] The communication device 1200 may be used to implement any of the methods described in Figures 7 to 10. Optionally, the processor 1210 may execute any of the methods described in Figures 7 to 10 based on the model.

[0466] In a possible embodiment, the communication device 1200 can implement the method performed by the second network element involved in any one of Figures 7 to 9.

[0467] In a possible embodiment, the communication device 1200 can implement the method performed by the first network element involved in any one of Figures 7 to 9.

[0468] In a possible embodiment, the communication device 1200 can implement the method performed by the second network element involved in Figure 10.

[0469] In a possible embodiment, the communication device 1200 can implement the method performed by the first network element involved in Figure 10.

[0470] Optionally, the processor 1210 is used to implement the functions of the above-mentioned processing module 1110, and the interface circuit 1220 is used to implement the functions of the above-mentioned transceiver module 1120.

[0471] When the communication device is a chip used in a network element, the network element chip implements the functions of the network element in the above method embodiments. The network element chip receives information from other modules in the network element (such as a radio frequency module or antenna), and the information is sent by other network elements to the network element; or the network element chip sends information to other modules in the network element (such as a radio frequency module or antenna), and the information is sent by the network element to other network elements.

[0472] It is understood that the processor involved in the various embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor. In addition, the memory involved in the various embodiments of the present application may include volatile memory, such as random access memory (RAM). The memory may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD) or solid state drive (SSD).

[0473] An embodiment of the present application provides another example of a communication device, which includes at least one processor and at least one memory, the at least one processor and the at least one memory being coupled, the at least one memory being used to store instructions, and when the instructions are executed by the at least one processor, the communication device executes the method in the above embodiment. Taking the communication device including a processor and a memory as an example, as shown in Figure 13, the communication device 1300 includes a processor 1310 and a memory 1320. The processor 1310 and the memory 1320 are coupled, and instructions are stored in the memory 1320. When the instructions stored in the memory 1320 are executed by the processor 1310, the communication device 1300 executes the method embodiment involved in any of Figures 7 to 10 above. Optionally, the communication device can also implement the function of any of the second network elements described above, or the function of any of the first network elements described above.

[0474] Optionally, the processor 1310 may deploy one or more models, such as the first model or the second model mentioned above. For example, if the processor includes an AI chip, the one or more models may be deployed on the AI ​​chip. Optionally, the processor 1310 may execute any of the method embodiments shown in Figures 7 to 10 above based on the one or more models.

[0475] In a possible embodiment, the communication device 1300 can implement the method performed by the second network element involved in any one of Figures 7 to 9.

[0476] In a possible embodiment, the communication device 1300 can implement the method performed by the first network element involved in any one of Figures 7 to 9.

[0477] In a possible embodiment, the communication device 1300 can implement the method performed by the second network element involved in FIG. 10 .

[0478] In a possible embodiment, the communication device 1300 can implement the method performed by the first network element involved in Figure 10.

[0479] An embodiment of the present application provides a communication system, which includes at least one network element and a second network element.

[0480] In one possible embodiment, the at least one network element includes, for example, a first network element, and any one of the at least one network element may implement, for example, any of the method embodiments performed by the first network element involved in Figures 7 to 9 above. The second network element may implement, for example, any of the method embodiments performed by the second network element involved in Figures 7 to 9 above.

[0481] In one possible embodiment, the at least one network element includes, for example, a first network element, and any one of the at least one network element may implement, for example, the method embodiment performed by the first network element involved in FIG10 . The second network element may implement, for example, the method embodiment performed by the second network element involved in FIG10 .

[0482] An embodiment of the present application provides a chip system, comprising: a processor and an interface, wherein the processor is configured to call and execute instructions from the interface, and when the processor executes the instructions, any of the method embodiments described in FIG. 7 to FIG. 10 is implemented.

[0483] An embodiment of the present application provides a computer-readable storage medium, which is used to store computer programs or instructions. When the computer-readable storage medium is executed, it implements any one of the method embodiments involved in Figures 7 to 10 above.

[0484] An embodiment of the present application provides a computer program product including instructions, which, when executed on a computer, implements any one of the method embodiments described in FIG. 7 to FIG. 10 .

[0485] The method steps in each embodiment of the present application can be implemented in hardware or in software instructions that can be executed by a processor. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, registers, hard disk, mobile hard disk, CD-ROM or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. The storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC. In addition, the ASIC can be located in a base station or a terminal. The processor and storage medium can also exist in a base station or a terminal as discrete components.

[0486] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are performed in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user device, or other programmable device. The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video disk; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or nonvolatile storage medium, or may include both volatile and nonvolatile types of storage media.

[0487] In the various embodiments of the present application, unless otherwise specified or there is any logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced by each other. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationships.

[0488] It should be understood that the various numbers used in the various embodiments of this application are merely for ease of description and are not intended to limit the scope of the embodiments of this application. The order of the sequence numbers of the above-mentioned processes does not necessarily imply a specific order of execution; the order of execution of the processes should be determined by their functions and inherent logic.

Claims

1. A communication method, characterized in that: Applied to a communication system including a first network element and a second network element, the method includes: The first network element receives first information from the second network element, where the first information indicates a time for determining inference data, where the inference data is used for inference based on a first model, where the first model is a local model of the first network element, and the first model and one or more models jointly participate in a task corresponding to vertical federated learning inference; The first network element determines, based on first data and the first model, an inference result of the first model, wherein the first data is inference data determined based on the first information; The first network element sends second information to the second network element, where the second information indicates a reasoning result of the first model; The second network element determines a first inference result based on the second information and third information, where the third information indicates an inference result of a second model, and the second model is one of the one or more models.

2. The method according to claim 1, characterized in that The first information is used to determine a time when the one or more models perform the inference data used by the task, and to determine a time when the first model performs the inference data used by the task.

3. The method according to claim 1 or 2, characterized in that The method further comprises: The second network element obtains an inference result of the second model based on second data and the second model, wherein the second data is determined based on the time for determining the inference data, and the second model is a local model of the second network element; The second network element determines the third information based on the inference result of the second model.

4. The method according to claim 1 or 2, characterized in that The method further comprises: The second network element receives the third information from a third network element, and the second model is a local model of the third network element.

5. The method according to any one of claims 1 to 4, characterized in that The first network element is a participant in vertical federated learning model reasoning, and / or the second network element is a vertical federated learning server.

6. The method according to any one of claims 1 to 5, characterized in that The first information indicates a time for determining the inference data, including: The first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates a moment when the first data is collected or generated, and the first time window is a time window for collecting or generating the first data.

7. The method according to claim 6, characterized in that The information of the first timestamp indicates the time when the inference data is collected or generated, including: The information of the first timestamp indicates a first moment and a time offset; The first moment is the moment when the inference data is started to be collected or generated, and the time offset represents the maximum deviation of the moment when the inference data is started to be collected or generated relative to the first moment; or The first moment is the moment when the collection or generation of the inference data is finished, and the time offset represents the maximum deviation of the moment when the collection or generation of the inference data is finished relative to the first moment.

8. The method according to claim 6, characterized in that The start time of the first time window is the time when the collection or generation of the inference data starts, and the end time of the first time window is the time when the collection or generation of the inference data ends.

9. The method according to any one of claims 6 to 8, characterized in that: The information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; The end time of the first time window; or A first period, where the first period is a period of multiple time windows included in the first time window.

10. The method according to any one of claims 1 to 9, characterized in that The first information further indicates at least one of the following: The moment when the first network element performs model inference; A first condition is used to trigger the first network element to perform model inference; or First indication information, where the first indication information is used to instruct the first network element to start model inference.

11. The method according to any one of claims 1 to 9, characterized in that The first information indicates a time for determining the inference data, including: The first information includes one or more of information on the moment when the first network element performs model inference, information on a first condition, or first indication information, and the one or more indications are used to determine the time of the inference data, wherein the first condition is used to trigger the first network element to perform model inference, and the first indication information is used to instruct the first network element to start model inference.

12. The method according to any one of claims 1 to 11, characterized in that The method further comprises: The second network element receives the following information sent by the first network element: First time information, the first time information indicating the time when the first data is actually collected or generated; and / or, The inference result of the first model includes sequence information of multiple inference results.

13. The method according to any one of claims 1 to 12, characterized in that The method further comprises: receiving information of at least one data from the first network element, where the at least one data is determined based on the first information, and the at least one data includes the first data; Send fourth information to the first network element, where the fourth information instructs the first network element to use the first data as inference data of the first model.

14. The method according to claim 13, characterized in that The information of the first data includes at least one of the following: Information on the serial number of the first data; Data characteristic information, the data characteristic information indicating an attribute of the first data; first time information, where the first time information indicates the time when the first data is actually collected or generated; first location information, where the first location information indicates a location where the first data is collected; or Access type information, where the access type information indicates an access type corresponding to the first data.

15. The method according to claim 14, characterized in that The first data satisfies at least one of the following: The difference between the number of the first data and the number of the inference data corresponding to the one or more models is less than or equal to a first threshold; The difference between the actual acquisition or generation time of the first data and the time of the inference data corresponding to the one or more models is less than or equal to a second threshold; The distance between the location where the first data is collected and the location of the inference data corresponding to the one or more models is less than or equal to a third threshold; or, The access type corresponding to the first data is the same as the access type corresponding to the inference data corresponding to the one or more models.

16. A communication method, characterized in that: Applied to a second network element, the method includes: Sending first information, where the first information indicates a time for determining inference data, where the inference data is used for inference based on a first model, where the first model is a local model of a first network element, and the first model and one or more models jointly participate in a task corresponding to vertical federated learning inference; receiving second information indicating a reasoning result based on the first model; A first inference result is determined based on the second information and third information, where the third information indicates an inference result of a second model, the second model being one of the one or more models.

17. The method according to claim 16, characterized in that The first information is used to determine a time when the one or more models perform the inference data used by the task, and to determine a time when the first model performs the inference data used by the task.

18. The method according to claim 16 or 17, characterized in that The method further comprises: Obtaining an inference result of the second model based on second data and a second model, wherein the second data is determined based on the time for determining the inference data; The third information is obtained according to the inference result of the second model.

19. The method according to claim 16 or 17, characterized in that The method further comprises: The third information is received from a third network element, where the second model is a local model of the third network element.

20. The method according to any one of claims 16 to 19, characterized in that: The first network element is a participant in vertical federated learning model reasoning, and / or the second network element is a vertical federated learning server.

21. The method according to any one of claims 16 to 20, characterized in that The first information indicates a time for determining the inference data, including: The first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates a moment when the inference data is collected or generated, and the first time window is a time window for collecting or generating the inference data.

22. The method according to claim 21, characterized in that The information of the first timestamp indicates the time when the inference data is collected or generated, including: The information of the first timestamp indicates a first moment and a time offset; The first moment is the moment when the inference data is started to be collected or generated, and the time offset represents the maximum deviation of the moment when the inference data is started to be collected or generated relative to the first moment; or The first moment is the moment when the collection or generation of the inference data is finished, and the time offset represents the maximum deviation of the moment when the collection or generation of the inference data is finished relative to the first moment.

23. The method according to claim 21, characterized in that The start time of the first time window is the time when the collection or generation of the inference data starts, and the end time of the first time window is the time when the collection or generation of the inference data ends.

24. The method according to any one of claims 21 to 23, characterized in that The information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; The end time of the first time window; or A first period, where the first period is a period of multiple time windows included in the first time window.

25. The method according to any one of claims 16 to 24, characterized in that: The first information further indicates at least one of the following: The moment when the first network element performs model inference; A first condition is used to trigger the first network element to perform model inference; or First indication information, wherein the first indication information is used to instruct the first network element to start model inference.

26. The method according to any one of claims 16 to 25, characterized in that The first information indicates a time for determining the inference data, including: The first information includes one or more of information on the moment when the first network element performs model inference, information on a first condition, or first indication information, and the one or more indications are used to determine the time of the inference data, wherein the first condition is used to trigger the first network element to perform model inference, and the first indication information is used to instruct the first network element to start model inference.

27. The method according to any one of claims 16 to 26, characterized in that The method further comprises: Receive the following information: First time information, where the first time information indicates the time when first data is actually collected or generated, where the first data is inference data of the first model; and / or, The inference result of the first model includes sequence information of multiple inference results.

28. The method according to any one of claims 16 to 27, characterized in that The method further comprises: receiving information of at least one data from a first network element, where the at least one data is determined based on the first information, and the at least one data includes first data; Send fourth information to the first network element, where the fourth information instructs the first network element to use the first data as inference data of the first model.

29. The method according to claim 28, characterized in that The information of the first data includes at least one of the following: Information on the serial number of the first data; Data characteristic information, the data characteristic information indicating an attribute of the first data; first time information, where the first time information indicates the time when the first data is actually collected or generated; first location information, where the first location information indicates a location where the first data is collected; or Access type information, where the access type information indicates an access type corresponding to the first data.

30. The method according to claim 29, wherein The first data satisfies at least one of the following: The difference between the number of the first data and the number of the inference data corresponding to the one or more models is less than or equal to a first threshold; The difference between the actual acquisition or generation time of the first data and the time of the inference data corresponding to the one or more models is less than or equal to a second threshold; The distance between the location where the first data is collected and the location of the inference data corresponding to the one or more models is less than or equal to a third threshold; or, The access type corresponding to the first data is the same as the access type corresponding to the inference data corresponding to the one or more models.

31. A communication method, characterized in that: The method comprises: receiving first information indicating a time for determining inference data, the inference data being used for inference based on a first model, the first model being a local model of a first network element, the first model and one or more models jointly participating in a task corresponding to vertical federated learning inference; Obtaining an inference result of a first model based on first data and the first model, where the first data is inference data determined based on the first information; Sending second information to the second network element, where the second information indicates the inference result of the first model.

32. The method according to claim 31, wherein The first network element is a participant in vertical federated learning model reasoning, and / or the second network element is a vertical federated learning server.

33. The method according to claim 31 or 32, characterized in that The first information indicates a time for determining the inference data, including: The first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates a moment when the reasoning data is collected or generated, and the first time window is a time window for collecting or generating the reasoning data.

34. The method according to claim 33, wherein The information of the first timestamp indicates the time when the inference data is collected or generated, including: The information of the first timestamp indicates the first moment and the time offset; wherein, the first moment is the moment when the collection or generation of reasoning data starts, and the time offset represents the maximum deviation of the moment when the collection or generation of reasoning data starts relative to the first moment; or, the first moment is the moment when the collection or generation of reasoning data ends, and the time offset represents the maximum deviation of the moment when the collection or generation of reasoning data ends relative to the first moment.

35. The method according to claim 33, wherein The start time of the first time window is the start time of collecting or generating the inference data, and the end time of the first time window is the end time of collecting or generating the inference data.

36. The method according to any one of claims 33 to 35, characterized in that The information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; The end time of the first time window; or The first cycle is the cycle of the first time window.

37. The method according to any one of claims 31 to 36, characterized in that The first information further includes at least one of the following: The moment when the first network element performs model inference; A first condition is used to trigger the first network element to perform reasoning based on the model; or The first indication information is used to instruct the first network element to start model inference.

38. The method according to any one of claims 31 to 37, characterized in that The first information indication is used to determine the time of inference data, including: the first information includes one or more of the information of the moment when the first network element performs model inference, the information of the first condition or the first indication information, and the one or more indications are used to determine the time of inference data.

39. The method according to any one of claims 31 to 37, wherein: The method further includes: sending the following information to the second network element: First time information, the first time information indicating the time when first data is actually collected or generated, the first data being inference data of the first model; and / or, The inference result of the first model includes sequence information of multiple inference results.

40. The method according to any one of claims 31 to 39, characterized in that The method further comprises: Sending information of at least one data to the second network element, where the at least one data is determined according to the first information, and the at least one data includes the first data; Fourth information is received, the fourth information indicating that the first data is used for inference.

41. The method according to claim 40, wherein The information of the first data includes at least one of the following: Information on the serial number of the first data; Data characteristic information, where the data characteristic information indicates an attribute of the first data; first time information, where the first time information indicates the time when the first data is actually collected or generated; first location information, where the first location information indicates a location where the first data is collected; or Access type information, where the access type information indicates an access type corresponding to the first data.

42. A communication method, characterized in that: The method comprises: First information is sent, where the first information indicates a time for determining training data, where the training data is used to train a first model, where the first model is associated with one or more models.

43. The method according to claim 42, characterized in that The method comprises: receiving second information sent from a first network element, where the second information indicates a training result of a first model of the first network element; A first training result is determined based on the training result of the first model and third information, where the third information indicates a training result of a second model, and the second model belongs to the one or more models.

44. The method according to claim 43, wherein The method further comprises: determining a training result of the second model based on second data and the second model to obtain the third information, wherein the second data is determined based on the time for determining the training data; or The third information is received.

45. The method according to any one of claims 42 to 44, characterized in that The first information indicates information for determining a first timestamp and / or information for a first time window, wherein the information for the first timestamp and / or information for the first time window indicates time for determining training data, wherein the first timestamp indicates a moment of collecting or generating the first data, the first time window is a time window for collecting or generating the first data, and the first data is training data for the first model.

46. ​​The method according to claim 45, characterized in that The information of the first timestamp indicates the time when the first data is collected or generated, including: The information of the first timestamp indicates a first moment and a time offset; wherein the first moment is the moment when the first data is started to be collected or generated, and the time offset is the maximum deviation of the moment when the first data is started to be collected or generated relative to the first moment; or, The first moment is the moment when the acquisition or generation of the first data is finished, and the time offset is the maximum deviation between the moment when the acquisition or generation of the first data is finished and the first moment.

47. The method according to claim 45, wherein The start time of the first time window is the time when the collection or generation of the first data starts, and the end time of the first time window is the time when the collection or generation of the first data ends.

48. The method according to any one of claims 45 to 47, characterized in that The information of the first time window includes at least one of the following: the length of the first time window; the start time of the first time window; The end time of the first time window; or The first cycle is the cycle of the first time window.

49. The method according to any one of claims 42 to 48, wherein: The first information further includes at least one of the following: The moment when the first network element performs model training; A first condition is used to trigger the first network element to perform training based on the model; or First indication information, the first indication information is used to instruct the first network element to start model training; The first model is a local model of the first network element.

50. The method according to any one of claims 42 to 49, characterized in that The first information indicates a time for determining training data, including: The first information includes one or more of information about the time when the first network element performs model training, information about the first condition, or first indication information, and the one or more information are used to determine the time of the training data.

51. The method according to any one of claims 42 to 50, characterized in that The method further comprises: Send one or more of the following information to the first network element: Information about the time when the first data is actually collected or generated; a time window in which first data is actually collected or generated, where the first data is training data for the first model; or The first training result includes sequence information of multiple training results.

52. The method according to any one of claims 42 to 51, characterized in that The method further comprises: receiving information of at least one data from a first network element, where the at least one data is determined based on a time for determining training data, the at least one data including first data; Fourth information is sent, where the fourth information indicates using first data for training, where the first data is training data for the first model.

53. The method according to claim 52, characterized in that The information of the first data includes at least one of the following: the serial number of the first data; Data characteristic information, the data characteristic information indicating characteristics of the first data; first time information, where the first time information indicates the time when the first data is actually collected or generated; First location information, the first location information indicates the location where the first data is collected; or Access type information, where the access type information indicates an access type corresponding to the first data; The first data is training data of the first model.

54. The method according to claim 52 or 53, characterized in that The first data satisfies at least one of the following conditions: The difference between the number of the first data and the number of the training data corresponding to the one or more models is less than or equal to a first threshold; The difference between the actual collection or generation time of the first data and the time of the training data corresponding to the one or more models is less than or equal to a second threshold; The distance between the location where the first data is collected and the location of the training data corresponding to the one or more models is less than or equal to a third threshold; or, The access type corresponding to the first data is the same as the access type corresponding to the training data corresponding to the one or more models; The first data is used to determine the training result of the first model.

55. The method according to any one of claims 42 to 54, wherein: The method further comprises: receiving fifth information indicating the first training result; According to the first training result, it is determined whether to end the training of the multiple models or to continue the training of the multiple models.

56. A communication method, characterized in that: The method comprises: receiving first information indicating a time for determining training data for training a first model; Obtaining a training result of a first model based on first data and a first model, wherein the first data is training data determined according to the first information, and the first model belongs to a plurality of models; Second information is sent to a second network element, where the second information indicates a training result of the first model, wherein the first model is associated with one or more models.

57. The method according to claim 56, characterized in that The first information indicates a time for determining training data, including: The first information includes information for determining a first timestamp and / or information for a first time window, wherein the first timestamp indicates a moment when the training data is collected or generated, and the first time window is a time window for collecting or generating the training data.

58. The method according to claim 57, wherein The information of the first timestamp indicates the time when the training data is collected or generated, including: The information of the first timestamp indicates the first moment and the time offset; wherein the first moment is the moment when the collection or generation of training data starts, and the time offset indicates the maximum deviation of the moment when the collection or generation of training data starts relative to the first moment; or, the first moment is the moment when the collection or generation of training data ends, and the time offset indicates the maximum deviation of the moment when the collection or generation of training data ends relative to the first moment.

59. The method according to claim 57 or 58, characterized in that The start time of the first time window is the start time of collecting or generating training data, and the end time of the first time window is the end time of collecting or generating training data.

60. The method according to any one of claims 57 to 59, wherein: The information of the first time window indicates at least one of the following: the length of the first time window; the start time of the first time window; The end time of the first time window; or The first cycle is the cycle of the first time window.

61. The method according to any one of claims 57 to 59, wherein: The first information further includes at least one of the following: The moment when the first network element performs model training; A first condition is used to trigger the first network element to perform training based on the model; or First indication information, the first indication information is used to instruct the first network element to start model training; The first model is a local model of the first network element.

62. The method according to any one of claims 56 to 61, characterized in that The first information indicates a time for determining training data, including: One or more of the information on the time when the first network element performs model training, the information on the first condition, or the first indication information, and the one or more indications are used to determine the time of the training data.

63. The method according to any one of claims 56 to 62, wherein: Send the following information to the second network element: First time information, the first time information indicating the time when the first data is actually collected or generated; and / or, The training result of the first model includes sequence information of multiple training results.

64. The method according to any one of claims 56 to 63, wherein: The method further comprises: Send information of at least one data to the second network element, where the at least one data is determined based on the first information and the at least one data includes the first data; and receive fourth information, where the fourth information indicates using the first data for training.

65. The method according to any one of claims 56 to 64, characterized in that The information of the first data includes at least one of the following: Information on the serial number of the first data; data characteristic information, the data characteristic information indicating an attribute of the first data; First time information, the first time information indicating the time when the first data is actually collected or generated; First location information, the first location information indicates a location where the first data is collected; or Access type information, where the access type information indicates an access type corresponding to the first data.

66. A communication system, characterized in that The communication system is configured to execute the method according to any one of claims 1 to 15.

67. A communication device, characterized in that include: A module for executing the method according to any one of claims 16 to 30; A module for executing the method according to any one of claims 31 to 41; A module for performing the method according to any one of claims 42 to 55; or A module for executing the method according to any one of claims 56 to 65.

68. A communication device, characterized in that The invention comprises a processor and an interface circuit, wherein the interface circuit is used to receive signals from other communication devices outside the communication device and transmit them to the processor or send signals from the processor to other communication devices outside the communication device, and the processor is used to implement the method according to any one of claims 1 to 15 through a logic circuit or execute code instructions, or to implement the method according to any one of claims 16 to 30, or to implement the method according to any one of claims 31 to 41, or to implement the method according to any one of claims 42 to 55, or to implement the method according to any one of claims 56 to 65.

69. A computer program product comprising instructions, characterized in that When the instruction is executed, it implements the method according to any one of claims 1 to 15, or implements the method according to any one of claims 16 to 30, or is used to implement the method according to any one of claims 31 to 41, or is used to implement the method according to any one of claims 42 to 55, or is used to implement the method according to any one of claims 56 to 65.

70. A computer-readable storage medium, characterized in that The storage medium stores a computer program or instruction. When the computer program or instruction is executed by the communication device, it implements the method according to any one of claims 1 to 15, or implements the method according to any one of claims 16 to 30, or is used to implement the method according to any one of claims 31 to 41, or is used to implement the method according to any one of claims 42 to 55, or is used to implement the method according to any one of claims 56 to 65.

Images