How to train a machine learning model

Abstract

According to one embodiment, a method for training a machine learning model is described, the method comprising the steps of: determining, for each of one or more federated learning clients, whether training data of the federated learning client satisfies a predetermined similarity criterion to test data of the federated learning server; and, for each of the clients, if the training data of the federated learning client satisfies the predetermined similarity criterion, updating a first version of the machine learning model running on the federated learning server to a second version of the machine learning model using an update generated by the federated learning client using the training data, and otherwise updating the training data of the federated learning client or the test data of the federated learning server. Figures related to the abstract: Figure 12

Description

[Technical Field]

[0001] The present disclosure relates to a method for training a machine learning model. [Background technology]

[0002] In communication networks, such as 5G mobile communication networks, it is important to be able to maintain a certain quality of service. To this end, network statistical analysis and predictive information, which can be generated from information such as load, resource usage, available components, component status, etc., may be monitored so that, for example, if overload is imminent, action can be taken to avoid degradation of service quality. Predictive information may be provided using machine learning models trained for this purpose. Therefore, it is desirable to efficiently train machine learning models in the context of communication systems, particularly to provide network analysis (including predictions). Summary of the Invention

[0003] According to one embodiment, a method for training a machine learning model is provided, the method comprising: determining, for each of one or more federated learning (FL) clients, whether training data of the federated learning client satisfies a predetermined similarity criterion to test data of a federated learning server; and, if, for each of the clients, the training data of the federated learning client satisfies the predetermined similarity criterion, updating a first version of the machine learning model running on the federated learning server to a second version of the machine learning model using updates generated by the federated learning client using the training data. [Brief explanation of the drawings]

[0004] In the drawings, like references generally refer to the same parts throughout the various views. The drawings are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the invention. In the following description, various aspects are described with reference to the following drawings: [Figure 1] FIG. 1 shows a mobile radio system. [Figure 2] Figure 2 shows the NWDAF AnLF (Analytics Logical Function) and the NWDAF MTLF (Model Training Logical Function). [Figure 3] Figure 3 shows a flow diagram illustrating the process for NWDAF service consumers to obtain analytical information from NWDAF (AnLF). [Figure 4] Figure 4 is a flow diagram illustrating federated learning (FL) in a 5G system. [Figure 5] FIG. 5 illustrates an approach for dealing with data heterogeneity in federated learning according to an embodiment. [Figure 6] FIG. 6 is a flow diagram illustrating the initialization and data collection phase of federated learning according to an embodiment. [Figure 7] FIG. 7 is a flow diagram illustrating the training phase of associative learning according to an embodiment. [Figure 8] FIG. 8 illustrates an implementation of federated learning according to an embodiment. [Figure 9] FIG. 9 illustrates an implementation of federated learning according to a further embodiment. [Figure 10] FIG. 10 illustrates an implementation of federated learning according to a further embodiment. [Figure 11] FIG. 11 illustrates a variation of the approach for dealing with data heterogeneity in the federated learning of FIG. 5 according to an embodiment. [Figure 12] FIG. 12 shows a flow diagram illustrating a method for training a machine learning model according to an embodiment. [Figure 13] FIG. 13 illustrates a data processing system according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0005] The following detailed description refers to the accompanying drawings, which show, by way of example, specific details and aspects of the present disclosure in which the invention may be practiced. Other aspects may be utilized, and structural, logical, and electrical changes may be made, without departing from the scope of the present disclosure. The various aspects of the present disclosure are not necessarily mutually exclusive, as some aspects of the present disclosure may be combined with one or more other aspects of the present disclosure to form new aspects.

[0006] Various examples corresponding to aspects of the present disclosure are described below:

[0007] Example 1 is a method for training a machine learning model as described above.

[0008] Example 2 includes the method of Example 1, further including a step in which the federated learning server generates a second version of the machine learning model by integrating updates from each of one or more federated learning clients where the test data of the federated learning server and the training data of the federated learning clients satisfy a predetermined similarity criterion.

[0009] Example 3 is the method of example 1 or 2, further comprising, for each of the one or more federated learning clients: sending information about the characteristics of the training data of the federated learning client to a federated learning server, and the federated learning server determining whether the test data of the federated learning server and the training data of the federated learning client satisfy a predetermined similarity criterion; or The method includes sending information about the characteristics of the test data of the federated learning server to the federated learning client, and the federated learning client determining whether the test data of the federated learning server and the training data of the federated learning client satisfy a predetermined similarity criterion.

[0010] Example 4 is the method of any one of Examples 1 to 3, wherein for each of the one or more federated learning clients: The predetermined similarity criterion is that each of one or more characteristics of the test data of the federated learning server is within a predetermined range of each of the corresponding characteristics of the training data of the federated learning client.

[0011] Example 5 is the method of Example 4, wherein the test data of the federated learning server and the training data of the federated learning client each include a plurality of data elements, each data element having a value for each of a plurality of components, the one or more characteristics of the test data of the federated learning server include a mean, a quantile, and / or a variance for each of at least some of the components of the data elements of the test data, and the one or more characteristics of the training data include a mean, a quantile, and / or a variance for each of at least some of the components of the data elements of the training data.

[0012] Example 6, in any one of the methods of Examples 1-5, includes a step of transmitting the machine learning model from a federated learning server to one or more federated learning clients, and each of the one or more federated learning clients deciding to update at least when the test data of the federated learning server and the training data of the federated learning client satisfy a predetermined similarity criterion.

[0013] Example 7 is the method of example 6, further comprising: determining, for each of the one or more federated learning clients, an updated version of the machine learning model according to the updates determined by the federated learning clients; comparing the accuracy of the updated version of the machine learning model on the test data of the federated learning client with the accuracy of the first version of the machine learning model on the test data of the federated learning server; The method includes updating the first version of the machine learning model to a second version of the machine learning model using an update by the federated learning client if the accuracy of the updated version of the machine learning model on the test data of the federated learning client is higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server.

[0014] Example 8 is the method of example 7, further comprising: for each of the federated learning clients, sending to the federated learning server an indication of the accuracy of the updated version of the machine learning model on the test data of the federated learning client; the federated learning server compares the accuracy of the updated version of the machine learning model on the federated learning client's test data with the accuracy of the first version of the machine learning model on the federated learning server's test data; or sending an indication of accuracy of the first version of the machine learning model on the test data from the federated learning server to the federated learning client; The method includes a step in which the federated learning client compares the accuracy of the updated version of the machine learning model on the test data of the federated learning client with the accuracy of the first version of the machine learning model on the test data of the federated learning server.

[0015] Example 9 is a method of any one of Examples 7 or 8, including a step of triggering an update of the test data of the federated learning server if, for each of one or more clients, the accuracy of the updated version of the machine learning model on the test data of the federated learning client is higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server, but the test data of the federated learning server and the training data of the federated learning client do not satisfy a predetermined similarity criterion.

[0016] Example 10, in any one of the methods of Examples 7-9, includes a step of avoiding using an update generated by the federated learning client to update the first version of the machine learning model if, for each of one or more clients, the test data of the federated learning server and the training data of the federated learning client do not satisfy a predetermined similarity criterion and the accuracy of the updated version of the machine learning model with respect to the test data of the federated learning client is not higher than the accuracy of the first version of the machine learning model with respect to the test data of the federated learning server.

[0017] Example 11 is a method of any one of Examples 7-10, including a step of triggering an update of the training data of the federated learning client when, for each of one or more clients, the test data of the federated learning server and the training data of the federated learning client do not satisfy a predetermined similarity criterion and the accuracy of the updated version of the machine learning model with respect to the test data of the federated learning client is not higher than the accuracy of the first version of the machine learning model with respect to the test data of the federated learning server.

[0018] Example 12 includes, in any one of the methods of Examples 1-11, implementing one or more federated learning clients by one or more mobile terminals of a mobile wireless communication system, or implementing both one or more federated learning clients and federated learning servers by a core network function of the mobile wireless communication system.

[0019] Example 13 is a method of any one of Examples 1-12, wherein each of one or more federated learning clients generates updates by training a machine learning model on the federated learning client's training data to perform a communication network analysis task.

[0020] Example 14 is the method of example 13, wherein the communication network analysis task is predicting communication resource needs.

[0021] A fifteenth embodiment is a data processing system configured to perform the method of any one of the first to fourteenth embodiments.

[0022] It should be noted that one or more features of any of the embodiments described above (and below) may be combined with any of the other embodiments. In particular, an embodiment described in the context of a device is equally valid for a method, and vice versa.

[0023] According to further aspects, there is provided a computer program product and a computer readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform any of the methods of any of the above embodiments.

[0024] Various embodiments will now be described in more detail.

[0025] FIG. 1 shows a mobile radio network 100 configured in accordance with 5G (Fifth Generation) as defined by, for example, 3GPP (Third Generation Partnership Project).

[0026] The wireless communication system 100 includes mobile wireless terminal devices 102, such as user equipment (UE), nano equipment (NE), etc. The mobile wireless terminal devices 102, also called subscriber terminals, form the terminal side, while the other components of the mobile communication system 100 described below are part of the mobile wireless communication network side, i.e., parts of a mobile wireless communication network (e.g., a Public Land Mobile Network (PLMN)).

[0027] Additionally, the wireless communication system 100 includes a Radio Access Network (RAN) 103, which may include multiple radio access network nodes, i.e., base stations, configured to provide wireless access in accordance with 5G (fifth generation) radio access technology (5G New Radio). It should be noted that the mobile wireless communication system 100 may be configured in accordance with Long Term Evolution (LTE) or other wireless communication standards (e.g., non-3GPP access such as WiFi), although 5G is used herein as an example. Each radio access network node can provide wireless communication over the air interface to a mobile wireless terminal device 102. It should be noted that the radio access network 103 may include any number of radio access network nodes.

[0028] The mobile radio communication system 100 further comprises a core network (5GC) 119 including an Access and Mobility Management Function (AMF) 101 connected to the RAN 103, a Unified Data Management (UDM) 104, and a Network Slice Selection Function (NSSF) 105. In the following examples, the UDM may further comprise an actual UE subscriber database, known for example as a Unified Data Repository (UDR). The core network 119 further comprises an Authentication Server Function (AUSF) 114 and a Policy Control Function (PCF) 115.

[0029] The core network 119 may have multiple core network slices 106, 107, and for each core network slice 106, 107, an operator may create multiple core network slice instances (NSIs) 108, 109. For example, the core network 119 may include a first core network slice 106 with three core network slice instances (CNIs) 108 for providing enhanced mobile broadband (eMBB) and a second core network slice 107 with three core network slice instances (CNIs) 109 for providing vehicle-to-everything (V2X).

[0030] Typically, when a core network slice is deployed (i.e., created), the network functions (NFs) are instantiated or (if already instantiated) referenced to form a core network slice instance (NSI), and the network functions belonging to the core network slice instance are configured with a core network slice instance identification.

[0031] Specifically, in the illustrated example, each instance 108 of the first core network slice 106 includes a first Session Management Function (SMF) 110 and a first User Plane Function (UPF) 111, and each instance 109 of the second core network slice 107 includes a second Session Management Function (SMF) 112 and a second User Plane Function (UPF) 113. The SMFs 110 and 112 are for processing Protocol Data Unit (PDU) sessions, i.e., for creating, updating, and removing PDU sessions and managing session contexts with the User Plane Function (UPF).

[0032] The RAN 103 and the core network 119 constitute the network side of the mobile radio link, or in other words, a mobile radio communication network. The mobile radio communication network and the mobile terminals accessing the mobile radio communication network together form a mobile radio communication system.

[0033] The mobile wireless communication system 100 may further include an Operation, Administration, and Maintenance (OAM) function (or entity) 116, implemented, for example, by one or more OAM servers connected to the RAN 103 and the core network 119 (connections not shown for clarity). The OAM 116 may include a Management Data Analytics Service (MDAS). The MDAS may provide, for example, analytical reports on the load of a network slice instance. Various factors may affect the load of a network slice instance, such as the number of UEs accessing the network, the number of QoS flows, and resource utilization of NFs associated with the network slice instance.

[0034] The core network 118 also includes a Network Repository Function (NRF).

[0035] The core network 119 may further include a Network Data Analytics Function (NWDAF) 117. The NWDAF is responsible for providing network analysis and / or forecasting information upon request from a network function. For example, a network function may request specific analytical information regarding the load level of a particular network slice instance. Alternatively, the network function may utilize a subscription service to ensure that it is notified by the NWDAF when the load level of a network slice instance changes or reaches a certain threshold. The NWDAF 117 may interface with various network functions on the mobile communication network side, such as the AMF 101, the SMFs 110 and 112, and the PCF 115. For simplicity, only the interface between the NWDAF 117 and the AMF 101 is shown.

[0036] For example, NWDAF analytics should enable monitoring the number of UEs registered to a network slice instance and their observed service experience. In addition to OAM enforcing service level agreement (SLA) guarantees, the 5GC NF may take action based on NWDAF slice QoE analytics to prevent further degradation of service experience in the network slice instance.

[0037] The NSSF 105 or AMF 101 may determine when a load balancing decision is needed to address issues identified, for example, by processing analysis results (i.e., network analysis and / or prediction information) provided by the NWDAF 117. For example, if a network slice instance is detected or predicted to suffer service experience degradation, new UE registrations or PDU sessions may no longer be assigned to that network slice instance by triggering a network slice load balancing mechanism. For example, the NSSF 105, AMF 101, and / or OAM 116 may simultaneously subscribe to both slice load analysis and slice service experience from the NWDAF 117. One or more subscriptions to one or more S-NSSAIs and NSIs are possible.

[0038] To generate network analysis and / or forecasting information, the NWDAF 117 collects the necessary input data (e.g., to derive slice service experience analysis), i.e., information for analyzing the state of network slice instances. The NWDAF 117 can obtain this type of information by subscribing to appropriately notified network functions.

[0039] According to 3GPP Release 17 (Rel-17), the NWDAF 117 is decomposed into two functions:

[0040] FIG. 2, for example, shows an NWDAF Analytics Logical Function (AnLF) 201 and an NWDAF Model Training Logical Function (MTLF) 202 connected via an Nnwdaf interface.

[0041] The NWDAF 201 containing the analytical logic function, denoted as NWDAF(AnLF) or NWDAF-AnLF or simply AnLF, is capable of performing inferences, deriving analytical information, and publishing analytical services, namely Nnwdaf_AnalyticsSubscription or Nnwdaf_AnalyticsInfo.

[0042] The NWDAF 202, which includes model training logic functions, may be referred to as the NWDAF(MTLF) or NWDAF-MTLF or simply MTLF, and may train machine learning (ML) models and expose new training services (e.g., provide trained models).

[0043] The NWDAF (AnLF) 201 provides an analytical service that can be contacted (e.g., via an NNF interface) by any NF (acting as a service consumer) 203 (e.g., AMF or SMF) that will be provided with analytical information.

[0044] The request parameters that the NF 203 inputs to the NWDAF (AnLF) 201 are, for example, an Analytic ID and an S-NSSAI. The output from the NWDAF (AnLF) 201 is, for example, statistics and / or predictions of the Analytic ID (referred to herein as network analytical information).

[0045] Examples of analysis IDs are UE communication, UE mobility, UE behavior, user data congestion, network performance, etc.

[0046] FIG. 3 shows a flow diagram 300 illustrating a flow for an NWDAF service consumer 301 (e.g., corresponding to an NF 303) to obtain analytical information from an NWDAF (AnLF) 302 (e.g., corresponding to an NWDAF (AnLF) 201).

[0047] At 303, the NWDAF Service Consumer 301 subscribes to the analytics service (and may later unsubscribe as well). The NWDAF (AnLF) 302 acknowledges at 304.

[0048] In response to the subscription, the NWDAF (AnLF) 302 sends the analysis information to the NWDAF service consumer 301 using one or more notification messages at 305 .

[0049] Analytical information, such as that provided by NWDAF (AnLF) 201, can be determined using, for example, machine learning models that may be trained to predict communications resource needs. Such machine learning models may be trained using federated learning.

[0050] Federated learning (FL) is a decentralized machine learning (ML) technique that trains ML models across multiple (FL) clients using local training datasets under the control of a central server (called the FL server).

[0051] It typically involves multiple training iterations, with each training iteration The FL server selects a certain number of clients to participate in this iteration of FL training. The FL server sends the current version of the model to the client (starting with the first iteration using the initial version of the model). Each FL client (also called an FL node) trains the model locally (i.e., updates the current version of the model, e.g., updates the weights of a neural network). Each FL client uses its own training data (also called local training data). The FL server pools the training results from the clients and aggregates them to generate a new version of the global model.

[0052] The resulting version (i.e., the new version of the last iteration) can be used by a model consumer such as the Network for Data Analysis and Ranging (AnLF)201, which can use it to make predictions from data obtained from the surroundings (such as communication resource needs).

[0053] In a 5G communication system, federated learning may be realized within a 5GC, for example, to provide NWDAF analysis, where the NWDAF (MTLF) is the FL server, several NWDAFs (MTLF) are FL clients, and the NWDAF (AnLF) is the consumer of the trained ML model (as in the example above), or it may be on top of the 5GC at the application layer for vertical, where the application in the Application Function (AF) is the FL server and the application in the UE is the FL client.

[0054] Live data in the communication network may be used by FL clients to train machine learning (ML) models.

[0055] FIG. 4 shows a flow diagram 400 illustrating federated learning in a 5G system, where the FL server is a first NWDAF 402 and the client is a second NWDAF 403.

[0056] At 406, a service consumer 401 (e.g., NWDAF (AnLF)) initiates training of an ML model by sending a subscription request to the FL server 402.

[0057] At 407, the FL server 402 provides FL parameters to the FL client 403 (eg, parameters that allow the client 403 to know the type of training data required and the initial version of the model).

[0058] At 408, each FL client 403 collects training data, for example, from an NRF 404 or some kind of NF 405 that provides relevant data.

[0059] Then, multiple iterations 409 are performed, each iteration including: Each client 403 performs local training (e.g., determines model updates for its respective “local” model, and each FL client sets the initial version of the model in the first iteration) and reports the model updates to the FL server 402 at 410. The FL server 402 aggregates model updates in 411. The FL server 402 can inform the service consumer 401 about the training status at 412. The service consumer 401 can provide an indication at 413 such as that the training should be ended or continued. The FL server 402 determines at 414 whether to continue or terminate the FL training service (e.g., according to some criteria, such as whether a desired accuracy has been achieved, or according to a client command). · (If training is to continue) The FL server 402 distributes the specification of the aggregate model (also called the "global" model) to the FL client 403 at 415. Each FL client 403 sets its local model to the global model (which will be used as the basis for training in the next iteration).

[0060] A key assumption of FL is that all clients train the same model using similar data. For example, in speech recognition, all clients may train an English model, or in UE mobility analysis, all MTLFs may train a model for urban UEs. Therefore, if there is data heterogeneity—i.e., if an FL client's local training data set is completely different from the local training data of other clients—training performance suffers. For example, in speech recognition, a client may have acquired data from German-speaking people as its local training data, or in UE mobility analysis, an MTLF may have acquired data from rural UEs (rather than urban) as its local training data. This typically leads to degradation of the global model's accuracy due to harmful updates from such clients (clients with "ill-fitting" local training data). Data heterogeneity can arise, for example, from temporal differences between clients' local data sets, geographic differences between clients, faulty behavior, malware, or data poisoning attacks.

[0061] It should be noted that even FL servers may have some "bad" data to use as test data, and therefore, FL servers may not be able to properly determine through testing whether a client has bad data (i.e., provided harmful updates). Furthermore, determining whether a client has bad local training data based on the accuracy of the client's local model compared to other clients' models may not work well because the difference between local model accuracy may be high even if there are small differences in local training data between clients due to partially trained models during training.

[0062] FIG. 5 illustrates an approach for dealing with data heterogeneity in federated learning according to an embodiment.

[0063] According to the approach shown in Figure 5: 1) In addition to details of the current global model, the FL server 501 sends meta data to the FL client 502 (according to various embodiments, the meta data includes statistical information about the FL server's 501 test data, denoted as G data statistics, and / or the accuracy of the FL server's 501 global model on these (“global”) test data, denoted as G accuracy). 2) In addition to updating the model, each FL client 502 attempts to detect data heterogeneity. 3) In addition to submitting model updates, each FL client 502 provides feedback to the FL server 501 regarding data heterogeneity. 4) In addition to model aggregation, the FL server 501 processes FL client feedback regarding data heterogeneity.

[0064] A detailed example of the approach of FIG. 5 is now described with reference to FIGS.

[0065] FIG. 6 shows a flow diagram 600 illustrating the initialization and data collection phase of federated learning according to an embodiment.

[0066] This flow is: FL Server 601, FL clients (only the i-th FL client 602 is depicted and discussed below); a global data store 603 containing global test data (i.e., an ADRF (Analysis Data Repository Function) containing data for each communication system); and The communication takes place between the local data stores of each FL client (only the local data store 604 of the i-th FL client 602 is depicted and discussed below).

[0067] An operation performed on (or by) the i-th FL client 602 and the local data store 604 of the i-th FL client 602 may be performed on (or by) each of multiple FL clients.

[0068] In an initialization step 605, the FL server 601 initializes the FL client 602 (eg, by an initialization message from the FL server 601 confirmed by an acknowledgement message by the FL client 602).

[0069] In a data collection step 606, the FL server 601 retrieves global test data from a global data store 603, and the client 602 retrieves local training data from a local data store 604 (via corresponding request and response messages).

[0070] FIG. 7 shows a flow diagram 700 illustrating the training phase of associative learning according to an embodiment.

[0071] Similar to the flow of FIG. 6, which the flow of FIG. 7 may follow, the flow of FIG. 7 takes place between an FL server 701, FL clients (only the i-th FL client 702 is shown here and will be discussed below), a global data store 703 containing global test data, and each FL client's local data store (only the i-th FL client 702's local data store 704 is shown here and will be discussed below).

[0072] The following abbreviations are used below: G-Model: Global model (current model version of FL server) L-Model: Local model of the FL client G-Data: Global test data (i.e., test data of the FL server) L-Data: Local training data G-Data statistics: Statistics about global test data L-Data statistics: Statistics about the local training data G-Accuracy: Accuracy of the G-Model relative to G data L-Accuracy: Accuracy of the L-Model for local test data (FL client test data)

[0073] Again, in the flow of FIG. 6, the operations performed on (or by) the i-th FL client 702 and the local data store 704 of the i-th FL client 702 may be performed on (or by) each of multiple FL clients.

[0074] In the 705, the FL server 701 supports the G-Model as well as G-Data statistics, and Accuracy of G-Model relative to G-Data is sent to the FL client 702.

[0075] At 706, the FL client 702 trains a local model using the L-Data and calculates the accuracy of the local model on the local test data.

[0076] Here, the FL client 702 compares the statistics of its local training data to the statistics of the global data, compares L-Accuracy to G-Accuracy, and one of the following is performed: ● If the FL client 702 detects that the local training data and the global test data are statistically different and that updating the model would degrade accuracy (i.e., L - Accuracy < G - Accuracy), then at 707, the client 702 does not participate in training (by correspondingly notifying the FL server 701) and obtains a new set of local training data from its local data store 704. ● If the FL client 702 detects that the local training data and the global test data are statistically different and that updating the model would improve accuracy (i.e., L - Accuracy > G - Accuracy), then the FL client concludes that its local training data is similar to the local training data of other clients and, at 708, requests the FL server 701 to update its global test data. The FL server 701 can then correspondingly update the global test data. ● If the FL client 702 detects that the local training data and the global test data are not statistically different, then at 709, it sends a model update to the FL server 701.

[0077] At 710, the FL server 701 aggregates the local models, i.e., determines a new version of the global model according to the received updates (if any).

[0078] FIG. 8 shows an implementation where the FL server 801 is implemented by MTLF, the global data store 803 is implemented by any NF or AF that provides global test data, the FL client 802 is implemented by MTLFs, and the local data store 804 is implemented by any NF or AF that provides local training data and local test data.

[0079] FIG. 9 shows an implementation where the FL server 901 is implemented by an MTLF, the global data store 903 is implemented by an ADRF providing global test data, the FL client 802 is implemented by MTLFs, and the local data store 804 is implemented by ADRFs providing local training and test data.

[0080] FIG. 10 shows an implementation where the FL server 1001 is implemented by an application in the AF, the global data store 1003 is implemented by any data source providing global test data, each FL client 1002 is implemented by an application running on the respective UE, and the local data store 1004 is implemented by any data source providing local training data and local test data.

[0081] FIG. 11 illustrates a variation of the approach for dealing with data heterogeneity in the federated learning of FIG. 5 according to an embodiment.

[0082] The approach of FIG. 11 differs from that of FIG. 5 in that data heterogeneity detection is performed by the FL server 1101 rather than the FL client 1102 .

[0083] To this end, each FL client 1102 sends the L-Data statistics and L-Accuracy indications to the FL server 1101 in addition to the determined model updates.

[0084] The FL server 1101 can then determine, among other things, for each FL client 1102 whether or not it should participate in the training (similar to 707 in FIG. 5) and notify the FL client 1102 accordingly.

[0085] Thus, the data heterogeneity detection logic may reside in the FL client 502, 1102, the FL server 501, 1101, or both.

[0086] In summary, according to various embodiments, a method is provided as shown in FIG.

[0087] FIG. 12 shows a flow diagram 1200 illustrating a method for training a machine learning model.

[0088] At 1201, for each of one or more federated learning clients, it is determined whether the test data of the federated learning server and the training data of the federated learning client meet a predetermined similarity criterion.

[0089] At 1202, for each of one or more clients, if the test data of the federated learning server and the training data of the federated learning client meet a predetermined similarity criterion, a first version of the machine learning model running on the federated learning server is updated to a second version of the machine learning model using an update generated by the federated learning client using the training data.

[0090] In 1203, for example, if the test data of the federated learning server and the training data of the federated learning client do not meet a predetermined similarity criterion, then, for example, the training data (and associated test data) of the federated learning client or the test data of the federated learning server is updated depending on whether the accuracy of an updated version of the machine learning model (in response to the update generated by the federated learning client) on the test data of the federated learning client is higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server.

[0091] According to various embodiments, in other words, if the test data used by the FL server and the training data used by the FL client are sufficiently similar, such as when the statistics of the test data used by the FL server match the statistics of the training data used by the FL client, the model is updated using updates from the FL client. Depending on whether the FL server or the FL client performs the check (i.e., whether the similarity criteria are met), there is a corresponding feedback mechanism between the FL client and the FL server, such as notification that the update determined by the FL client will not (or should not) be used to update the model running on the FL server. Furthermore, information for performing the check (and possibly further checks, such as the accuracy check described above) is exchanged between the FL client and the FL server. In particular, according to various embodiments, G-Accuracy, G-Data statistics, and information not involving the FL client are communicated, as described above. Furthermore, the training data and / or the test data may be updated (or refreshed), particularly depending on the results of the check.

[0092] It should be noted that, from the perspective of the federated learning client, updating a first version of the machine learning model to a second version of the machine learning model may mean that the federated learning client sends the update that it generated to the federated learning server. From the perspective of the federated learning server, updating a first version of the machine learning model to a second version of the machine learning model may include aggregating the updates generated by the federated learning client.

[0093] The test data of the federated learning server and the training data of the federated learning client may meet a predetermined similarity criterion if and only if their statistics (with respect to statistics) meet the predetermined similarity criterion, for example.

[0094] It should be further noted that a new “version” (and alternative synonyms thereof, such as “release”) of a machine learning model to which an existing machine learning model is updated corresponds to a new, modified model obtained through local updates by one or more federated learning clients, aggregation by a federated learning server, or some modification of the existing model, i.e., an update may include modification of parameters such as (neural network) weights.

[0095] The similarity between the training data set of one or more FL clients and the test data set of the FL server can be measured without sharing the actual data set among the federated learning server and one or more FL clients; rather, it can be achieved by sending some characteristics of the test data, such as statistics, distribution, data sparsity, or other data similarity measures, from the federated learning server to each of the one or more federated learning clients, or by sending some characteristics of the training data, such as statistics, distribution, data sparsity, or other data similarity measures, from the federated learning clients to the federated learning server.

[0096] As in the example above, multiple iterations of the process described with reference to FIG. 12 can be performed to train a machine learning model.

[0097] The approach of FIG. 12 particularly allows achieving high accuracy of the trained model so that the trained model performs well in real applications.

[0098] The ML model trained by the FL according to the approach of Figure 12 may be used for decision making in various situations, for example, PCF, UPF, SMF, AMF, and other NFs, and also AFs can use the NWDAF analysis provided by such a model for resource planning.

[0099] Inaccuracies in ML models lead to incorrect predictions and therefore incorrect decisions, ultimately resulting in poor performance. For example, an inaccurate model for estimating UE mobility leads to incorrect predictions of the number of UEs, thus leading to incorrect resource planning and ultimately to poor UE performance. The approach of Figure 12 allows for improving the accuracy model trained by FL, thereby achieving more accurate predictions and therefore better decisions, and therefore improved performance.

[0100] The method of FIG. 12 may be performed, for example, by a data processing system such as that shown in FIG.

[0101] FIG. 13 illustrates a data processing system according to an embodiment.

[0102] The data processing system includes one or more data processing devices 1300 (e.g., mobile terminals (UEs) and / or server computers), each including a communication interface 1301 (e.g., for FL client-FL server communication), a memory 1302 (e.g., for storing local models, global models, or model updates, and program code), and a processing unit 1303 (e.g., a CPU) for performing various functions of the method.

[0103] Components of a data processing device are realized by one or more circuits. A "circuit" may be understood as any kind of logic implementation entity, which may be a dedicated circuit or a processor running software stored in memory, firmware, or any combination thereof. Thus, a "circuit" may be a hardwired logic circuit or a programmable logic circuit, e.g., a programmable processor such as a microprocessor. A "circuit" may also be a processor running software, e.g., any kind of computer program. Any other kind of implementation of each of the above-mentioned functions may also be understood as a "circuit."

[0104] While particular embodiments have been described, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the embodiments of the present disclosure as defined by the appended claims, the scope of which is accordingly indicated by the appended claims, and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced.

Claims

1. 1. A method executed by a data processing system for training a machine learning model, the method comprising: For each of the one or more federated learning clients, determining whether the training data of the federated learning client satisfies a predetermined similarity criterion to the test data of the federated learning server; and for each of the federated learning clients, if training data of the federated learning client satisfies the predetermined similarity criterion, updating a first version of the machine learning model running on the federated learning server to a second version of the machine learning model using an update generated by the federated learning client using the training data; A method comprising:

2. 2. The method of claim 1, further comprising: generating a second version of the machine learning model by integrating updates of each of the one or more federated learning clients whose test data and training data satisfy the predetermined similarity criteria; A method comprising:

3. 10. The method of claim 1, wherein for each of the one or more federated learning clients: sending information about characteristics of the training data of the federated learning client to the federated learning server, and the federated learning server determining whether the training data of the federated learning client satisfies the predetermined similarity criterion; or a step of transmitting information about characteristics of test data from the federated learning server to the federated learning client, the federated learning client determining whether the training data of the federated learning client satisfies the predetermined similarity criterion.

4. 10. The method of claim 1, wherein for each of the one or more federated learning clients: The method, wherein the predetermined similarity criterion is that each of one or more characteristics of the test data of the federated learning server is within a predetermined range of a corresponding characteristic of the training data of the federated learning client.

5. 5. The method of claim 4, wherein the test data of the federated learning server and the training data of the federated learning client each include a plurality of data elements, each data element having a value for each of a plurality of components, the one or more characteristics of the test data of the federated learning server include a mean, quartiles, and / or variance for each of at least some of the components of the data elements of the test data, and the one or more characteristics of the training data include a mean, quartiles, and / or variance for each of at least some of the components of the data elements of the training data.

6. 2. The method of claim 1, further comprising: transmitting the machine learning model from the federated learning server to the one or more federated learning clients; and each of the one or more federated learning clients determining the update at least if the training data of the federated learning client satisfies the predetermined similarity criterion.

7. 10. The method of claim 6, further comprising: determining, for each of the one or more federated learning clients, an updated version of the machine learning model according to updates determined by the federated learning client; comparing the accuracy of the updated version of the machine learning model on test data at the federated learning client to the accuracy of the first version of the machine learning model on test data at the federated learning server; updating the first version of the machine learning model to a second version of the machine learning model using an update by the federated learning client when accuracy of the updated version of the machine learning model on test data at the federated learning client is higher than accuracy of the first version of the machine learning model on test data at the federated learning server; A method comprising:

8. 10. The method of claim 7, further comprising: transmitting, for each federated learning client, an indication of accuracy of the updated version of the machine learning model on the test data of the federated learning client to the federated learning server; the federated learning server compares the accuracy of the updated version of the machine learning model on the federated learning client's test data with the accuracy of the first version of the machine learning model on the federated learning server's test data; or sending an indication of accuracy of the first version of the machine learning model on test data from the federated learning server to the federated learning client; The method includes a step in which the federated learning client compares the accuracy of an updated version of the machine learning model on test data at the federated learning client with the accuracy of a first version of the machine learning model on test data at the federated learning server.

9. 8. The method of claim 7, further comprising: for each of a plurality of federated learning clients, triggering an update of the test data of the federated learning server when the accuracy of the updated version of the machine learning model on the test data of the federated learning client is higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server, but the training data of the federated learning client does not satisfy the predetermined similarity criterion.

10. 8. The method of claim 7, comprising: for each of a plurality of federated learning clients, avoiding using an update generated by the federated learning client to update a first version of the machine learning model if the training data of the federated learning client does not satisfy the predetermined similarity criterion and the accuracy of the updated version of the machine learning model on the test data of the federated learning client is not higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server.

11. 8. The method of claim 7, further comprising: for each of one or more federated learning clients, triggering an update of the training data of the federated learning client if the training data of the federated learning client does not satisfy the predetermined similarity criterion and the accuracy of the updated version of the machine learning model on the test data of the federated learning client is not higher than the accuracy of the first version of the machine learning model on the test data of the federated learning server.

12. 2. The method of claim 1, comprising implementing the one or more federated learning clients by one or more mobile terminals of a mobile radio communication system, or implementing both the one or more federated learning clients and the federated learning server by a core network function of the mobile radio communication system.

13. 2. The method of claim 1, wherein each of the one or more federated learning clients generates the updates by training the machine learning model on training data of the federated learning client to perform a communication network analysis task.

14. 14. The method of claim 13, wherein the communication network analysis task is predicting communication resource needs.

15. A data processing system configured to perform the method of any one of claims 1 to 14.

Images