A model training method and device for network data analysis function based on federated learning

Abstract

The embodiment of the present invention provides a model training method and device for network data analysis function based on federated learning. The method comprises: obtaining a data sample set at time t and the time series information corresponding to the data sample set; determining a training set according to the data sample set, wherein the training set comprises at least one training task, wherein the training task comprises multiple support sets and multiple query sets, wherein the support set comprises multiple data samples, and the query set comprises multiple data samples; and training a meta-learning network according to the training set to determine the parameters of the meta-learning network. The scheme of the present invention can realize a local feature training optimization process with a very short training time for each round of iteration.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a model training method and device for a network data analysis function based on federated learning. Background Art

[0002] In the process of data sharing and distributed deep learning, there is often the problem of data privacy leakage. In order to solve this problem, the federated learning method is often used. In the application scenario of federated learning, the training data set is used on multiple participants (such as different organizations, different institutions, different devices, and different geographical locations). Users / organizations / alliances do not share their local training data sets, and the data of any party will not be leaked to other participants. Through joint federated learning, multiple parties with data can jointly train one or more models, which can improve the task performance of the participants' local models while ensuring that private data is not leaked, breaking down data silos. This method is called federated learning.

[0003] In the existing 5G technology solutions, the scenarios of artificial intelligence in communication networks are studied, and the needs and data that need to be analyzed are sorted out. NWDAF (NNetwork Data Analytics Function) uses network operation data collected from 5G network functions (NFs), terminal and network-related statistical data obtained from the operation and maintenance management system OAM (Operation Administration and Maintenance), and application data obtained from third-party applications through NEF.

[0004] The analysis results generated by NWDAF are also output to 5G network functions, OAM or third-party applications. 5G network functions, OAM or application layer servers can use the analysis results of NWDAF to perform different optimization operations.

[0005] The distributed intelligent architecture based on Federated Learning, as a distributed machine learning framework, allows participants to conduct joint modeling without sharing data, technically breaking information silos and realizing distributed intelligent collaboration, while also improving model training efficiency.

[0006] Meta-learning is a subfield of machine learning that applies automatic learning algorithms to the metadata of machine learning experiments. Today's AI systems can learn a complex skill from scratch with a lot of time and experience. However, if you want the agent to master multiple skills and adapt to multiple environments, you should not train each skill from scratch in each environment. Instead, the agent needs to learn how to learn multiple new tasks by reusing past experience. This method of learning how to learn, also known as meta-learning, is the only way to a multifaceted agent that can continuously learn multiple new tasks. Meta-learning no longer focuses on acquiring labels. Instead, it tries to make the machine learn faster from a small amount of data. If this technique is used, less data can be used for training than general models.

[0007] In the current federated learning between multiple NWDAF instances, the central server integrates the weights of different individuals. The participants in the federated learning framework of multiple NWDAF instances are relatively dispersed and distributed. The goal of federated learning is to minimize the empirical function so that the model can fit the data collected by several NWDAF participants. Therefore, the model parameters are currently used for weighted averaging (FedAvg) and then sent to each NWDAF participant / alliance. Each NWDAF participant uses the model for reasoning and the next round of calculations.

[0008] Since the purpose of NWDAF federated learning is to obtain a shared, averaged global model, all participants in NWDAF federated learning will eventually obtain the same model effect, so each NWDAF federated participant does not have its own local personalized features. However, due to the differences among the participants of each NWDAF in actual applications and the Non-IID distribution of data, in many cases, each local node of NWDAF federated learning participant needs to build personalized model technology, such as for personalized recommendation, personalized analysis and processing, etc. Therefore, it is necessary to implement personalized customization of the global model to improve the effect on independent nodes.

[0009] In addition, each NWDAF federation participant will continuously collect new real-time data, so the data generated by each participant will continue to accumulate and increase, and may reach a massive amount of data over time. However, the resources and computing power of the NWDAF federation participants in federated learning are limited, and in federated learning, the participants first calculate in parallel and then centrally aggregate by the central server. Therefore, even if the calculation time of a single participant is extended, it will affect and drag down the entire system. Therefore, if the current NWDAF federated learning system uses all the local massive data in actual applications to train a local personalized model, it will encounter serious performance, cost, and efficiency problems. Therefore, it is also necessary to solve the problem of how to construct a reliable personalized model with a small amount of data.

[0010] Taking the language model of the next input prediction word on the keyboard of a mobile phone terminal in NWDAF federated learning as an example, different customers may use different languages, regions, habits, and preferences. Therefore, the input prediction method may be to collect not only the words entered by the user so far, but also various other user information in real time, such as context features, which words the user often uses, what application they are currently using, if they are chatting with someone, they have sent messages to this person, current location, time, external factors and other data, which will gradually increase over time. So what needs to be solved is to use a lower cost to solve the problem of highly personalized and localized shared global models.

[0011] In addition, due to the "federalism" and "fairness" of each NWDAF federation in federated learning, each participant can join and exit the federated learning system at any time. When new NWDAF participants start to collect and provide very little data, it is also a problem how new NWDAF participants in federated learning can construct a reliable initial model with a small amount of data. Summary of the invention

[0012] The present invention provides a model training method and device for network data analysis function based on federated learning, which can automatically obtain and update information such as adjustment context content, and can also realize a local feature training optimization process with a very short iterative training time for each round.

[0013] To solve the above technical problems, the embodiments of the present invention provide the following solutions:

[0014] A model training method for network data analysis function based on federated learning, the method comprising:

[0015] Obtaining a data sample set at time t and time series information corresponding to the data sample set;

[0016] Determine a training set according to the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples;

[0017] The meta-learning network is trained according to the training set to determine the parameters of the meta-learning network.

[0018] Optionally, there are A data samples at time t, represented by |Di|=A;

[0019] Among them, D i Represents the data of the i-th network data analysis function NWDAF participant

[0020] Optionally, determining a training set according to the data sample set includes:

[0021] Repeatedly extract multiple tasks from the Di data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where NK training samples form the support set and NL test samples form the query set;

[0022] According to the time sequence, the latest preset number of tasks are randomly selected from Di, and the tasks include N*K support sets and N*L query sets.

[0023] Optionally, training a meta-learning network according to the training set to determine parameters of the meta-learning network includes:

[0024] According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network.

[0025] Optionally, generate model parameters of the meta-learning network include:

[0026] In the first step of sampling 1 training task m in ;

[0027] The parameters of the model network φ 0 Assign to the network of task m, and get

[0028] Using the support set of task m, the learning rate of task m is Perform 1 optimization and update

[0029] Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of

[0030] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 ;

[0031] Sample 1 task n and set the parameter φ 1 Assign to task n, and get

[0032] Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update

[0033] Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on

[0034] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ;

[0035] In training tasks Repeat the above process and after calculating all the tasks, we get

[0036] Optional model training methods for network data analysis functions based on federated learning also include:

[0037] Transmit the updated NWDAF Go to the NWDAF federated learning system central server;

[0038] The receiving central server executes the global aggregation formula The global θ obtained t ;S represents all NWDAF participating parties;

[0039] According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The process of obtaining the network parameters for this round of training.

[0040] An embodiment of the present invention further provides a model training device for a network data analysis function based on federated learning, the device comprising:

[0041] A context environment dynamic perception module is used to obtain a data sample set at time t and time series information corresponding to the data sample set; determine a training set based on the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples;

[0042] The meta-learning training module trains the meta-learning network according to the training set to determine the parameters of the meta-learning network.

[0043] Optionally, there are A data samples at time t, represented by |Di|=A;

[0044] Among them, D i Represents the data of the i-th network data analysis function NWDAF participant

[0045] Optionally, determining a training set according to the data sample set includes:

[0046] Repeatedly extract multiple tasks from the Di data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where NK training samples form the support set and NL test samples form the query set;

[0047] According to the time sequence, the latest preset number of tasks are randomly selected from Di, and the tasks include N*K support sets and N*L query sets.

[0048] Optionally, training a meta-learning network according to the training set to determine parameters of the meta-learning network includes:

[0049] According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network.

[0050] Optionally, generate model parameters of the meta-learning network include:

[0051] In the first step of sampling 1 training task m in ;

[0052] The parameters of the model network φ 0 Assign to the network of task m, and get

[0053] Using the support set of task m, the learning rate of task m is Perform 1 optimization and update

[0054] Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of

[0055] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 ;

[0056] Sample 1 task n and set the parameter φ 1 Assign to task n, and get

[0057] Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update

[0058] Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on

[0059] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ;

[0060] In training tasks Repeat the above process and after calculating all the tasks, we get

[0061] Optionally, the model training device for the network data analysis function based on federated learning also includes:

[0062] Transceiver module, used to transmit updated NWDAF Go to the NWDAF federated learning system central server;

[0063] The receiving central server executes the global aggregation formula The global θ obtained t ;

[0064] According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The process of obtaining the network parameters for this round of training.

[0065] An embodiment of the present invention further provides a processor-readable storage medium, wherein the processor-readable storage medium stores processor-executable instructions, and the processor-executable instructions are used to enable the processor to execute the method as described above.

[0066] The above solution of the present invention includes at least the following beneficial effects:

[0067] The above scheme of the present invention obtains a data sample set at time t and the time series information corresponding to the data sample set; determines a training set based on the data sample set, wherein the training set includes at least one training task, wherein the training task includes multiple support sets and multiple query sets, wherein the support set includes multiple data samples, and the query set includes multiple data samples; trains a meta-learning network based on the training set to determine the parameters of the meta-learning network. Model training and tuning can be performed using only a small number of customer samples in each round of calculation of the NWDAF federated learning participant, and can automatically obtain and update information such as the adjustment context content, and can also achieve a local feature training optimization process with a very short iterative training time for each round. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A schematic diagram of a flow chart of a model training method for a network data analysis function based on federated learning according to an embodiment of the present invention;

[0069] Figure 2 In the embodiment of the present invention, Schematic diagram of

[0070] Figure 3 A schematic diagram of a meta-learning framework in an embodiment of the present invention;

[0071] Figure 4 This is a module diagram of a model training device with a network data analysis function based on federated learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0073] like Figure 1As shown, an embodiment of the present invention provides a model training method for a network data analysis function based on federated learning, the method comprising:

[0074] Step 11, obtaining a data sample set at time t and time series information corresponding to the data sample set;

[0075] Step 12, determining a training set according to the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples;

[0076] Step 13: train the meta-learning network according to the training set to determine the parameters of the meta-learning network.

[0077] This embodiment of the present invention can perform model training and tuning using only a small number of customer samples in each round of calculation by the NWDAF federated learning participants, and can automatically obtain and update information such as adjustment context content, and can also achieve a local feature training optimization process with a very short iterative training time for each round.

[0078] In an optional embodiment of the present invention, at time t, there are A (including newly generated data at time t and existing data before time t) data samples, expressed as |Di|=A;

[0079] Assume that S represents all NWDAF participants, D i Represents the data of the i-th network data analysis function NWDAF participant

[0080] In an optional embodiment of the present invention, step 12 may include:

[0081] Step 121, repeatedly extract multiple tasks from the Di data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where NK training samples form a support set, and NL test samples form a query set;

[0082] Step 122, randomly select the latest preset number of tasks from Di according to the time sequence, and the tasks include a support set of N*K and a query set of N*L.

[0083] In specific implementation, the context environment dynamic perception module mainly completes the following tasks: maintaining the cache of data samples before time Di t and the time series related information corresponding to the samples, and obtaining data samples at time t and the time series related information corresponding to the samples.

[0084] The context environment dynamic perception module is mainly used to complete the real-time perception of the current environment of the participants, and prepare the tasks required for training for the meta-learning module according to actual needs.

[0085] 1. If Di is just added to the federated learning system, the context environment dynamic perception module repeatedly extracts several tasks from the Di data. Each task includes N*K (N*K) N-ways, K-shot data samples and N*L (N*L) N-ways, L-shot data samples. Generally speaking, we call these NK training samples the support set (Support Set) and the NL test samples the query set (Query Set). These tasks (Train Task) are represented as use Figure 2 shown.

[0086] If NWDAF participant i is a federated system that has joined NWDAF, i will perform a round of local training at regular intervals. The context environment dynamic perception module randomly selects the latest few tasks (including N*K support sets and N*L query sets) from Di according to the time sequence, which can be called Test Tasks, as the training sets for each round of regular training for the second part of the meta-learning training module.

[0087] In an optional embodiment of the present invention, step 13 may include:

[0088] According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network.

[0089] In an optional embodiment of the present invention, the model parameters of the meta-learning network are generated include:

[0090] In the first step of sampling 1 training task m in ;

[0091] The parameters of the model network φ 0 Assign to the network of task m, and get

[0092] Using the support set of task m, the learning rate of task m is Perform 1 optimization and update

[0093] Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of

[0094] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 ;

[0095] Sample 1 task n and set the parameter φ 1 Assign to task n, and get

[0096] Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update

[0097] Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on

[0098] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ;

[0099] In training tasks Repeat the above process and after calculating all the tasks, we get

[0100] Optional model training methods for network data analysis functions based on federated learning also include:

[0101] Transmit the updated NWDAF Go to the NWDAF federated learning system central server;

[0102] The receiving central server executes the global aggregation formula The global θ obtained t ;S represents all NWDAF participating parties;

[0103] According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The process of obtaining the network parameters for this round of training.

[0104] like Figure 3As shown, when the method is implemented specifically, the meta-learning training module implements a model-agnostic meta-learning framework (taking a central server and one participant as an example, in reality it is generally a central server + multiple participants).

[0105] Since federated learning usually minimizes the following objective function: Each participant generates model parameters locally After that, it is uploaded to the central server, which aggregates the model parameters of each participant to obtain It will be sent to all participants as the next model delivery content.

[0106] When NWDAF participant i just joins the federated learning system, the first local optimization pre-training process is as follows:

[0107] 1. In the first step of sampling 1 training task m in (or several training tasks in 1 batch). Set the parameters φ of the model network 0 Assign to the network of task m, and get (Initial ).

[0108] 2. Use the Support Set of task m, based on the learning rate of task m Perform 1 optimization and update

[0109] 3. Based on 1 optimization Use Query Set to calculate the loss function loss of task m: And calculate right gradient.

[0110] 4. Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 .

[0111] 5. Sample 1 task n and set the parameter φ 1 Assign to task n, get (Initial ).

[0112] 6. Then use the training data of task n and the learning rate α based on task n n ,right Perform an optimization and update

[0113] 7. Based on 1 optimization Use Query Set to calculate the loss function loss of task n: And calculate The gradient on

[0114] 8. Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 .

[0115] 9. In the training task Repeat steps 1-8 to calculate all tasks.

[0116] 10. Transmit the updated NWDAF To the central server of the NWDAF federated learning system. The central server then executes the global aggregation formula To achieve the global θ t .

[0117] 11. The central server of the NWDAF federated learning system sends the aggregated θ t To each participating location.

[0118] When NWDAF participant i is started, it executes the local optimized federated learning training process at regular intervals:

[0119] 12. Each NWDAF participant receives the θ broadcast from the central server. t After that, the context environment dynamic perception module extracts a small amount of data and arranges the newer TestTask in time sequence. TestTask can be used as the training Repeat the previous steps 1 to 9 to obtain the network parameters for this round of training.

[0120] 13. After that, each round of local training calculation only needs to repeat step 12.

[0121] Therefore, except for the data required by each NWDAF participant when it is first started, which is the same as the traditional federated learning, the subsequent federated learning training process at regular intervals (set by the federated learning system) only requires a small amount of the latest data. This can not only meet the architectural requirements of NWDAF federated learning, but also allow users to use multiple data sets distributed in different locations to train machine learning models, while preventing data leakage and complying with strict data privacy laws, so as to personalize the contextual information of each NWDAF participant. In addition, since the number of tasks that need to be selected in each round of training is very small, the amount of computation and cost are greatly reduced, thereby greatly improving the performance and efficiency of NWDAF federated learning.

[0122] The present invention aims at the problem that the models that are finally distributed to each federation of NWDAF in the existing NWDAF federation learning do not have local model personalization, and the limitations and defects of performance and efficiency brought about by localization, and proposes a new method and architecture of federation learning and meta-learning by combining NWDAF. The introduction of meta-learning training module and context environment dynamic perception module can be used for training and tuning with a small amount of data, which can not only improve the personalization, performance and efficiency of each NWDAF participant model, but also merge the contextual time-series information of each participant, which is suitable for the conditions and situations where the resources and computing power of each NWDAF participant are limited. The design of the functional architecture of the context environment dynamic perception module for dynamic perception, adjustment, analysis and maintenance of the environmental context. The design, function and architecture of the NWDAF meta-learning training module based on federated learning. It not only meets the main architecture requirements of NWDAF federation learning, but also realizes the prediction personalization and localization optimization of the participants of NWDAF federation learning. It can also use a small amount of NWDAF data for training and tuning to suit the conditions and situations where the resources and computing power of each NWDAF participant are extremely limited, and it basically does not affect the effect of the trained model.

[0123] like Figure 4 As shown, an embodiment of the present invention further provides a model training device 40 for network data analysis function based on federated learning, the device comprising:

[0124] The context dynamic perception module 41 is used to obtain a data sample set at time t and the time series information corresponding to the data sample set; determine a training set according to the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples;

[0125] The meta-learning training module 42 trains the meta-learning network according to the training set to determine the parameters of the meta-learning network.

[0126] Optionally, there are A data samples at time t, represented by |Di|=A;

[0127] Among them, D i Represents the data of the i-th network data analysis function NWDAF participant

[0128] Optionally, determining a training set according to the data sample set includes:

[0129] Repeatedly extract multiple tasks from the Di data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where NK training samples form the support set and NL test samples form the query set;

[0130] According to the time sequence, the latest preset number of tasks are randomly selected from Di, and the tasks include N*K support sets and N*L query sets.

[0131] Optionally, training a meta-learning network according to the training set to determine parameters of the meta-learning network includes:

[0132] According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network.

[0133] Optionally, generate model parameters of the meta-learning network include:

[0134] In the first step of sampling 1 training task m in ;

[0135] The parameters of the model network φ 0 Assign to the network of task m, and get

[0136] Using the support set of task m, the learning rate of task m is Perform 1 optimization and update

[0137] Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of

[0138] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 ;

[0139] Sample 1 task n and set the parameter φ 1 Assign to task n, and get

[0140] Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update

[0141] Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on

[0142] Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ;

[0143] In training tasks Repeat the above process and after calculating all the tasks, we get

[0144] Optionally, the model training device for the network data analysis function based on federated learning also includes:

[0145] Transceiver module, used to transmit updated NWDAF Go to the NWDAF federated learning system central server;

[0146] The receiving central server executes the global aggregation formula The global θ obtained t ;

[0147] According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The network parameters of this round of training are obtained through the process.

[0148] It should be noted that the device is a device corresponding to the above method embodiment, and all implementation methods in the above method embodiment are applicable to the embodiment of the device and can achieve the same technical effect.

[0149] The embodiment of the present invention further provides a processor-readable storage medium, wherein the processor-readable storage medium stores processor-executable instructions, wherein the processor-executable instructions are used to cause the processor to execute the method described above. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0150] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0151] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0152] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0153] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0154] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0155] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks.

[0156] In addition, it should be noted that in the apparatus and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. Moreover, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order, and some steps can be performed in parallel or independently of each other. For those of ordinary skill in the art, it is understood that all or any steps or components of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or a network of computing devices in hardware, firmware, software or a combination thereof, which can be achieved by those of ordinary skill in the art using their basic programming skills after reading the description of the present invention.

[0157] Therefore, the purpose of the present invention can also be achieved by running a program or a group of programs on any computing device. The computing device can be a well-known general device. Therefore, the purpose of the present invention can also be achieved by simply providing a program product containing a program code that implements the method or device. That is to say, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any well-known storage medium or any storage medium developed in the future. It should also be pointed out that in the device and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. In addition, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order. Some steps can be performed in parallel or independently of each other.

[0158] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A model training method for network data analysis function based on federated learning, characterized in that: The method comprises: Obtaining a data sample set at time t and time series information corresponding to the data sample set; Determine a training set according to the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples; Training a meta-learning network according to the training set to determine parameters of the meta-learning network; Training a meta-learning network according to the training set to determine parameters of the meta-learning network includes: According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network. Generate model parameters of the meta-learning network include: sampling 1 training task m in ; The parameters of the model network φ 0 Assign to the network of task m, and get Using the support set of task m, the learning rate of task m is Perform 1 optimization and update Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of Use this gradient to multiply the learning rate α of the meta-learning network me ta, update φ 0 , we get φ 1 ; Sample 1 task n and set the parameter φ 1 Assign to task n, and get Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ; In training tasks Repeat the above process and after calculating all the tasks, we get The method further comprises: Update the NWDAF entry method Transmit to the NWDAF federated learning system central server; The receiving central server executes the global aggregation formula The global θ obtained t ;S represents all NWDAF participating parties; According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The model parameters for this round of training are obtained through the process of 2. The model training method for network data analysis function based on federated learning according to claim 1 is characterized in that: There are A data samples at time t, represented by |D i |=A; Among them, D i Represents the data of the i-th network data analysis function NWDAF participant 3. The model training method for network data analysis function based on federated learning according to claim 2 is characterized in that: Determining a training set according to the data sample set includes: Repeatedly in D i Extract multiple tasks from the data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where N*K training samples form a support set and N*L test samples form a query set; According to the time sequence, from D i A preset number of the latest tasks are randomly selected from the , wherein the tasks include a support set of N*K and a query set of N*L.

4. A model training device for network data analysis function based on federated learning, characterized in that: The device comprises: A context environment dynamic perception module is used to obtain a data sample set at time t and time series information corresponding to the data sample set; determine a training set based on the data sample set, wherein the training set includes at least one training task, the training task includes multiple support sets and multiple query sets, the support set includes multiple data samples, and the query set includes multiple data samples; A meta-learning training module, used to train the meta-learning network according to the training set and determine the parameters of the meta-learning network; Training a meta-learning network according to the training set to determine parameters of the meta-learning network includes: According to the training set, the first localized pre-training is performed on the meta-learning network to generate the model parameters of the meta-learning network. Generate model parameters of the meta-learning network include: sampling 1 training task m in ; The parameters of the model network φ 0 Assign to the network of task m, and get Using the support set of task m, the learning rate of task m is Perform 1 optimization and update Based on one optimization Use the query set to calculate the loss function loss of task m: And calculate right The gradient of Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 0 , we get φ 1 ; Sample 1 task n and set the parameter φ 1 Assign to task n, and get Using the training data of task n, the learning rate α based on task n n ,right Perform an optimization and update Based on one optimization Use the query set to calculate the loss function loss of task n: And calculate The gradient on Use this gradient to multiply the learning rate α of the meta-learning network meta , update φ 1 , we get φ 2 ; In training tasks Repeat the above process and after calculating all the tasks, we get The device also includes: Transceiver module, used to update NWDAF Transmit to the NWDAF federated learning system central server; The receiving central server executes the global aggregation formula The global θ obtained t , S represents all NWDAF participating parties; According to the global θ t , extract a preset number of data samples, and arrange the newer test tasks in time sequence, and treat the test tasks as the training tasks of this time Repeated The model parameters for this round of training are obtained through the process of 5. The model training device for network data analysis function based on federated learning according to claim 4, characterized in that: There are A data samples at time t, represented by |D i |=A; Among them, D i Represents the data of the i-th network data analysis function NWDAF participant 6. The model training device for network data analysis function based on federated learning according to claim 5, characterized in that: Determining a training set according to the data sample set includes: Repeatedly in D i Extract multiple tasks from the data, each task includes N*K (N*K) N-ways, K-shot data samples, N*L (N*L) N-ways, L-shot data samples, where N*K training samples form a support set and N*L test samples form a query set; According to the time sequence, from D i A preset number of the latest tasks are randomly selected from the , wherein the tasks include a support set of N*K and a query set of N*L.

7. A processor-readable storage medium, characterized in that: The processor-readable storage medium stores processor-executable instructions, and the processor-executable instructions are used to enable the processor to execute the method according to any one of claims 1 to 3.

Images