Graph data federal forgetting learning method and device based on memory perception and storage medium
By adopting the graph data federated forgetting learning method based on memory perception in the federated learning recommendation system, the problem of user data forgetting is solved, fast and effective user data culling and privacy protection are achieved, and the performance of the recommendation system is maintained.
Patent Information
- Application Number
- CN202510089879.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-04-25
AI Technical Summary
In a federated learning recommendation system with advanced expansion, when a user makes a forget request, it is difficult for the prior art to quickly and effectively remove the user's data contribution from the global model, resulting in loss of model performance and insufficient privacy protection.
The graph data federated forgetting learning method based on memory perception is used to determine the contribution of user data to the model through anonymous user feature clustering and distributed parameter importance weight calculation, and the model is pruned and fine-tuned on a third-party server to achieve the forgetting of user data.
It realizes that when a user makes a forget request, quickly calculates and eliminates the user's unique contribution to the model, avoids the need to retrain the model, effectively protects the user's data privacy, and maintains the performance of the recommendation system.
Smart Images
Figure CN119886392A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of distributed artificial intelligence and relates to a graph data federated forgetting learning method based on memory perception. Background Art
[0002] As the number of user personal terminals continues to increase, a large amount of valuable data is generated in many e-commerce platforms, social media platforms and other applications. The recommendation system built with this data has become an indispensable and important tool for the development of many companies. Users can also rely on the recommendation system to browse the parts they are interested in from massive data, which greatly enhances people's ability to interact with the real world. In order to avoid user data leakage and privacy issues caused by collecting data from a large number of users, centralized storage and application when building a recommendation system, the recommendation system integrates the federated learning framework to ensure that the user's interaction graph data is only stored on the local device. At the same time, with the introduction of some legislation in various countries, users are given the "right to be forgotten". Users can choose to delete their own data at any time to protect their privacy.
[0003] Currently, in the recommendation system of federated learning, researchers only focus on studying one aspect of federated recommendation or forgetting of recommendation system. For example, in federated recommendation, the federated learning frameworks FCF and FedMF are integrated into the recommendation system, and the first-order subgraph of user-interaction items is trained locally. However, since the first-order subgraph contains less information, it performs poorly in terms of recommendation accuracy. The latest federated recommendation framework FedPerGNN achieves high-order expansion of the user-interaction item subgraph in a privacy-preserving manner by designing a new federated learning mechanism with a third-party server, enabling anonymous interaction of user information, enriching local subgraph information, and effectively improving the accuracy of recommendations. In terms of recommendation forgetting, it is mainly divided into two types: approximate forgetting of graph data and precise forgetting. Both types can provide effective privacy protection for users participating in the recommendation system. However, there is no research combining these two aspects for the following reasons: First, in federated recommendation, such as the high-order information expansion mechanism of FedPerGNN, there will be an extension of forgotten user nodes in the local training data of most users. When a user makes a forget request, whether it is an accurate or approximate recommendation forgetting method, most users need to forget locally at the same time to eliminate the influence of forgotten nodes, which is impractical; second, the existing recommendation forgetting algorithms are all performed under the assumption of linear models, such as SGC, which cannot provide an effective forgetting method for nonlinear models. However, in most recommendation systems, nonlinear models occupy a dominant position and their model performance is better than linear models; finally, the recommendation system has a huge number of users and very little local storage data, while conventional federated forgetting methods, such as UPGA and Rapid Retraining, usually forget the client as a whole. When the amount of data in the client is too small, a poor model correction effect will be produced during forgetting, resulting in data being difficult to forget and the accuracy of the forgetting model is not high.
[0004] In order to fill the gap between federated recommendation and recommendation forgetting, the present invention designs a memory-aware graph data federated forgetting learning method based on the FedPerGNN framework. The core of this method is to realize clustering through anonymous user features on a third-party server, and perform distributed parameter importance weight calculations locally on the user. The central server aggregates the model's memory perception of this type of data, and then uses a parameter pruning process to eliminate the impact of this type of data on the model, and adds a fine-tuning process to restore the performance loss of the model due to pruning. Similar to the human brain generating electricity when stimulated by memory, the data that needs to be forgotten is input into the model, and the gradient values generated by each dimension of the model are observed to determine which dimensions of the model parameters are affected. Therefore, it is called a memory-aware method. Summary of the invention
[0005] The problem that the present invention solves is that in the context of a federated learning recommendation system with high-order extensions, when a user makes a forget request, the server can calculate the contribution of the user's locally stored interactive subgraph data to the global GNN (graph neural network) model through a memory-aware method, and then prune this part of the contribution on the global model to achieve forgetting of the user's data; at the same time, fine-tuning training should be performed locally using the data of the remaining users to maintain the performance of the recommendation system model.
[0006] The technical solution of the present invention is as follows:
[0007] A memory-aware graph data federated forgetting learning method includes a user anonymous clustering algorithm, a memory-aware model parameter importance weight calculation method, and a model pruning method; the specific steps are as follows:
[0008] Step 1: Cluster all local users of the federated learning recommendation system that participates in high-level information expansion, and save the clustering results on a third-party server;
[0009] Step 2: Select representative users from each user category, calculate the gradient data of all representative users in each dimension, and determine the specific dimensions of each user category that affect the global GNN model based on the gradient data, and calculate the influence matrix;
[0010] Step 3: Analyze the common features of the influence matrix of each user category from the frequency domain perspective through two-dimensional discrete Fourier decomposition; remove the common features from the influence matrix of the forgotten user, retain the unique features of the forgotten user, and construct an importance weight matrix;
[0011] Step 4: Prune the global GNN model on the central server according to the importance weight matrix to obtain a pruned model; users of the same category as the forgetting user fine-tune the pruned model, and the central server aggregates the fine-tuned models to obtain a final forgetting model.
[0012] Furthermore, in step 1, the specific steps of clustering all local users are as follows:
[0013] Step 1.1: After the federated learning training is completed, the central server sends the global GNN model and clustering notification to the local user;
[0014] Step 1.2: User U k After receiving the clustering notification, the local data is input into the global GNN model again, and the feature extraction layer of the global GNN model outputs the user embedding Interactive Item Embedding And neighbor node embedding Where k = 1, 2, ..., N;
[0015] Step 1.3: Locally Use local sensitive hashing to encrypt data and get h k , and h k Upload to a third-party server;
[0016] Step 1.4: Obtain the hash encryption value h = {h1,h2,h3,...,h N}, use Kmean++ algorithm to cluster users; record the user ID under each category, record the category-user ID information on the third-party server, and complete the clustering process.
[0017] Furthermore, in step 2, the specific steps of calculating the influence matrix are as follows:
[0018] Step 2.1: In each user category, randomly select the same number of users as representative users U r , where r = 1, 2, ..., N r , N r is the number of all representative users; the central server sends the global GNN model to each category of representative users;
[0019] Step 2.2: Each category represents user U r Its local map data D r Input global GNN model W = {ω ij}, the output H(D r ; W); where ω ij are model parameters, i and j represent the dimensions of the global GNN model W respectively;
[0020] Step 2.3: Calculate the representative user U r The gradient data of each dimension, which constitutes the influence matrix I r ; The calculation formula of the gradient data is:
[0021]
[0022] in, is the gradient data, representing user U r The local map data D r For the model parameter ω ij The impact of
[0023] Step 2.4: Calculate the influence matrix of all representative users and upload it to the central server.
[0024] Furthermore, in step 3, the specific steps of constructing the importance weight matrix are as follows:
[0025] Step 3.1: For any representative user U r Upload Impact Matrix I r Perform a two-dimensional discrete Fourier transform to obtain the frequency domain representation of the influence matrix F r (u,v), where u and v are coordinates in the frequency domain;
[0026] Step 3.2: Calculate F r (u,v) at frequency f j =Amplitude value under (u,v) Where, j = 1, 2, ..., u × v;
[0027] Step 3.3: The influence matrix of all representative users at frequency f j =(u,v) is constructed as a set, expressed as Assumptions Follows the Gamma distribution, that is in and They are Shape parameters and scale parameters of
[0028] Step 3.4: Estimate using variational Bayesian methods and The approximate posterior distribution of
[0029] Step 3.5: Take samples and obtain and The mean and Thus we get The distribution of As the representative value of the frequency amplitude value; after calculating the representative values of all frequencies, we get
[0030] Step 3.6: For the forgotten user U', calculate the influence matrix I U' , execute steps 3.1 to 3.5 to obtain I U' The frequency domain representation F U' And the amplitude value at each frequency |F U' |; from |F U' |Minus|F rep |, which means to remove each frequency in I U' The common features in the frequency domain of are used to obtain unique features;
[0031] Step 3.7: Using F U' The original phase ∠F U' And the amplitude difference |F obtained in step 3.6 U'|-|F rep |Construct a new frequency domain representation F U' -F rep , and convert it back to the spatial domain and name it I' U' ; to I' U' Normalize and get the global GNN model parameter importance weight matrix Ω U' .
[0032] Furthermore, the specific process of step 3.4 is as follows:
[0033] Step 3.4.1: Build joint probability distribution in Indicates that in a given and In the case of data The likelihood function of and is the prior distribution; specifically, let and Follows the Gamma distribution, that is Among them, a0, b0, c0, and d0 are hyperparameters set based on prior experience; and The prior distribution is Gamma distribution, so the variational distribution is set to in yes The variational distribution of is a Gamma distribution with shape parameters and scale parameters respectively. and yes The variational distribution of is also a Gamma distribution, with shape parameters and scale parameters respectively. and λ includes all variational parameters and
[0034] Step 3.4.2: Use the ELBO to find the parameter λ so that the variational distribution As close as possible to the true posterior distribution
[0035] Furthermore, in step 4, the model pruning and model performance recovery method includes the following specific steps:
[0036] Step 4.1: The central server calculates the parameter weight Ω based on the forgotten user U' Prune the global GNN model W, and the pruning process is expressed as in represents the Hadamard product;
[0037] Step 4.2: The central server will prune the model W * Send to other users of the same category as the forgotten user, and other users of the same category use the local graph interaction data to prune the model W * Fine-tune and upload the fine-tuned model to the central server;
[0038] Step 4.3: The central server aggregates the fine-tuned models to obtain the final forget model.
[0039] An electronic device comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor; when the processor executes the computer program, the electronic device executes the memory-aware graph data federated forgetting learning method.
[0040] A storage medium comprising a computer program, which, when executed on an electronic device, enables the electronic device to execute the memory-aware graph data federated forgetting learning method.
[0041] Beneficial effects of the present invention: The present invention can quickly determine which dimensions of the model are affected by the forgotten user through a memory-aware method; and uses two-dimensional discrete Fourier transform and variational Bayes method to calculate the unique influence of the forgotten user, based on which the model is pruned to achieve the forgetfulness of the target user. Through the method of the present invention, when a user makes a forget request in a federated learning recommendation system with high-order information, the unique contribution of the forgotten user to the model can be quickly calculated and eliminated without retraining the recommendation model, effectively protecting the user's data privacy. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is an algorithm flow chart of the method of the present invention;
[0043] Figure 2 is a flow chart of the clustering method of the present invention;
[0044] Figure 3 It is a detailed design diagram of the algorithm for calculating the importance weights of model parameters of the present invention;
[0045] Figure 4 It is a flow chart of model pruning and model performance recovery of the present invention. DETAILED DESCRIPTION
[0046] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation modes of the present invention will be described in detail below with reference to the accompanying drawings.
[0047] Federated learning combined with recommendation systems can improve the personalized effect of recommendation systems through multi-party data sharing and joint training under the premise of protecting user privacy, while avoiding the centralized storage of sensitive data to ensure data security and privacy. However, with the establishment of relevant legislation on the "right to be forgotten", when a user makes a forget request, the server cannot quickly remove the user's contribution from the trained recommendation model, and usually needs to retrain the model on the data of the remaining users, resulting in a waste of time and computing resources. For example, in the product recommendation system of an e-commerce platform, the e-commerce platform trains the recommendation model in a federated learning manner by accessing the user's local interaction data with the product. When a user asks the platform to delete the influence of his data from the model, the memory-aware graph data federated forgetting learning method can quickly locate the influence of the user's data on the model, and prune the model on the server side, thereby effectively removing the user's contribution; this not only avoids retraining the recommendation model, but also protects the user's legitimate rights and interests without affecting the accuracy of other users' recommendations. Figure 1 The algorithm flow chart is shown in FIG. 1 , and the method specifically includes the following steps:
[0048] Step 1: Cluster all local users of the federated learning recommendation system that participates in high-level information expansion, and save the clustering results on a third-party server, such as Figure 2 As shown, the specific process is as follows:
[0049] Step 1.1: After the federated learning training is completed, the central server sends the global GNN model and clustering notification to the local user;
[0050] Step 1.2: Any user U k After receiving the clustering notification, the local data is input into the global GNN model again, and the feature extraction layer of the global GNN model outputs the user embedding Interactive Item Embedding And neighbor node embedding Where k = 1, 2, ..., N;
[0051] Step 1.3: Locally Use local sensitive hashing to encrypt data and get h k , and h k Upload to a third-party server;
[0052] Step 1.4: Obtain the hash encryption value h = {h1,h2,h3,...,h N}, use Kmean++ algorithm to cluster users; record the user ID under each category, record the category-user ID information on the third-party server, and complete the clustering process.
[0053] Step 2: Select representative users from each user category, calculate the gradient data of all representative users in each dimension, and determine the specific dimensions of each user category that affect the global GNN model based on the gradient data, and calculate the influence matrix; the specific process is:
[0054] Step 2.1: In each user category, randomly select the same number of users as representative users U r , where r = 1, 2, ..., N r , N r is the number of all representative users; the central server sends the global GNN model to each category of representative users;
[0055] Step 2.2: Each category represents user U r Its local map data D r Input global GNN model W = {ω ij}, the output H(D r ; W); where ω ij are model parameters, i and j represent the dimensions of the global GNN model W respectively;
[0056] Step 2.3: Calculate the representative user U r The gradient data of each dimension, which constitutes the influence matrix I r ; The calculation formula of the gradient data is:
[0057]
[0058] in, is the gradient data, representing user U r Local map data D r For the model parameter ω ij The impact of
[0059] Step 2.4: Calculate the influence matrix of all representative users and upload it to the central server.
[0060] Step 3: Analyze the common features of the influence matrix of each user category from the frequency domain perspective through two-dimensional discrete Fourier decomposition; remove the common features from the influence matrix of the forgotten user, retain the unique features of the forgotten user, and construct an importance weight matrix; Figure 3 As shown, the specific process is:
[0061] Step 3.1: For any representative user U r Upload Impact Matrix I r Perform a two-dimensional discrete Fourier transform to obtain the frequency domain representation of the influence matrix F r (u,v), where u and v are coordinates in the frequency domain;
[0062] Step 3.2: Calculate F r (u,v) at frequency f j = amplitude under (u,v) Where, j = 1, 2, ..., u × v;
[0063] Step 3.3: The influence matrix of all representative users at frequency f j =(u,v) is constructed as a set, expressed as Assumptions Follows the Gamma distribution, that is in and They are Shape parameters and scale parameters of
[0064] Step 3.4: Estimate using variational Bayesian methods and The approximate posterior distribution of Specifically:
[0065] Step 3.4.1: Build joint probability distribution in Indicates that in a given and In the case of data The likelihood function of and is the prior distribution; specifically, let and Follows the Gamma distribution, that is Among them, a0, b0, c0, and d0 are hyperparameters set based on prior experience; and The prior distribution is Gamma distribution, so the variational distribution is set to in yes The variational distribution of is a Gamma distribution with shape parameters and scale parameters respectively. and yes The variational distribution of is also a Gamma distribution, with shape parameters and scale parameters respectively. and λ includes all variational parameters and
[0066] Step 3.4.2: Use the ELBO to find the parameter λ so that the variational distribution As close as possible to the true posterior distribution
[0067] Step 3.5: Take samples and obtain and The mean and Thus we get The distribution of As the representative value of the frequency amplitude value; after calculating the representative values of all frequencies, we get
[0068] Step 3.6: For the forgotten user U', execute step 2 to calculate the influence matrix I U' , execute steps 3.1 to 3.5 to obtain I U' The frequency domain representation F U' And the amplitude value at each frequency |F U' |; from |F U' |Minus|F rep |, which means to remove each frequency in I U' The common features in the frequency domain of are used to obtain unique features;
[0069] Step 3.7: Using F U' The original phase ∠F U' And the amplitude difference |F obtained in step 3.6 U' |-|F rep |Construct a new frequency domain representation F U' -F rep , and convert it back to the spatial domain and name it I' U' ; to I' U' Normalize and get the global GNN model parameter importance weight matrix Ω U' .
[0070] Step 4: Prune the global GNN model on the central server according to the importance weight matrix to obtain a pruned model, thereby eliminating the contribution of the category to which the forgotten user belongs to the model; users of the same category as the forgotten user fine-tune the pruned model, and the central server aggregates the fine-tuned model to obtain the final forgotten model, thereby recovering the performance loss of the model caused by pruning; Figure 4 As shown, the specific steps are as follows:
[0071] Step 4.1: The central server calculates the parameter weight Ω based on the forgotten user U' Prune the global GNN model W, and the pruning process is expressed as in represents the Hadamard product;
[0072] Step 4.2: The central server will prune the model W* Send to other users of the same category as the forgotten user, and other users of the same category use the local graph interaction data to prune the model W * Fine-tune and upload the fine-tuned model to the central server;
[0073] Step 4.3: The central server aggregates the fine-tuned models to obtain the final forget model.
[0074] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any form of limitation on the present invention. Those skilled in the art should fully understand that it is completely feasible to modify the technical solutions described in the above embodiments or to perform equivalent replacements on any part or all of the technical features therein. These modifications or replacements, as long as they do not deviate from the scope of protection determined by the claims of the present invention, should be regarded as reasonable extensions of the present invention.
Claims
1. A memory-aware graph data federated forgetting learning method, characterized in that: The following steps are involved: Step 1: Cluster all local users of the federated learning recommendation system that participates in high-level information expansion, and save the clustering results on a third-party server; Step 2: Select representative users from each user category, calculate the gradient data of all representative users in each dimension, and determine the specific dimensions of each user category that affect the global GNN model based on the gradient data, and calculate the influence matrix; Step 3: Analyze the common characteristics of the influence matrix of each user category from the frequency domain perspective through two-dimensional discrete Fourier decomposition; The common features are removed from the influence matrix of the forgotten user, the unique features of the forgotten user are retained, and an importance weight matrix is constructed; Step 4: Prune the global GNN model on the central server according to the importance weight matrix to obtain a pruned model; users of the same category as the forgetting user fine-tune the pruned model, and the central server aggregates the fine-tuned models to obtain a final forgetting model.
2. The memory-aware graph data federated forgetting learning method according to claim 1 is characterized in that: In step 1, the specific steps of clustering all local users are as follows: Step 1.1: After the federated learning training is completed, the central server sends the global GNN model and clustering notification to the local user; Step 1.2: User U k After receiving the clustering notification, the local data is input into the global GNN model again, and the feature extraction layer of the global GNN model outputs the user embedding Where k = 1, 2, ..., N; Step 1.3: Locally Use local sensitive hashing to encrypt data and get h k , and h k Upload to a third-party server; Step 1.4: Obtain the hash encryption value h of all users based on the third-party server. N }, use Kmean++ algorithm to cluster users; record the user ID under each category, record the category-user ID information on the third-party server, and complete the clustering process.
3. The memory-aware graph data federated forgetting learning method according to claim 1 or 2, characterized in that: In step 2, the specific steps of calculating the influence matrix are as follows: Step 2.1: In each user category, randomly select the same number of users as representative users U r , where r = 1, 2, ..., N r , N r is the number of all representative users; the central server sends the global GNN model to each category of representative users; Step 2.2: Each category represents user U r Its local map data D r Input global GNN model W = {ω ij }, the output H(D r ; W); where ω ij are model parameters, i and j represent the dimensions of the global GNN model W respectively; Step 2.3: Calculate the representative user U r The gradient data of each dimension, which constitutes the influence matrix i r ; The calculation formula of the gradient data is: in, is the gradient data, representing user U r The local map data D r For the model parameter ω ij The impact of Step 2.4: Calculate the influence matrix of all representative users and upload it to the central server.
4. The memory-aware graph data federated forgetting learning method according to claim 3 is characterized in that: In step 3, the specific steps of constructing the importance weight matrix are as follows: Step 3.1: For any representative user U r Upload Impact Matrix I r Perform a two-dimensional discrete Fourier transform to obtain the frequency domain representation of the influence matrix F r (u,v), where u and v are coordinates in the frequency domain; Step 3.2: Calculate F r (u,v) at frequency f j =Amplitude value under (u,v) Where, j = 1, 2, ..., u × v; Step 3.3: The influence matrix of all representative users at frequency f j =(u,v) is constructed as a set, expressed as set up Follows the Gamma distribution, that is in and They are Shape parameters and scale parameters of Step 3.4: Estimate using variational Bayesian methods and The approximate posterior distribution of Step 3.5: Take samples and obtain and The mean and Thus we get The distribution of As the representative value of the frequency amplitude value; after calculating the representative values of all frequencies, we get Step 3.6: For the forgotten user U', calculate the influence matrix I U' , and get I U' The frequency domain representation F U' And the amplitude value at each frequency |F U' |; from |F U' |Minus|F rep |, which means to remove each frequency in I U' The common features in the frequency domain of are used to obtain unique features; Step 3.7: Using F U' The original phase ∠F U' And the amplitude difference |F U' |-|F rep |Construct a new frequency domain representation F U' -F rep , and convert it back to the spatial domain and name it I' U' ; to I' U' Normalize and get the global GNN model parameter importance weight matrix Ω U' .
5. The memory-aware graph data federated forgetting learning method according to claim 4 is characterized in that: The specific process of step 3.4 is as follows: Step 3.4.1: Build joint probability distribution in Indicates that in a given and In the case of data The likelihood function of and is the prior distribution; specifically, let and Follows the Gamma distribution, that is Among them, a0, b0, c0, and d0 are hyperparameters set based on prior experience; and The prior distribution is Gamma distribution, so the variational distribution is set to in yes The variational distribution of is a Gamma distribution with shape parameters and scale parameters respectively. and yes The variational distribution of is also a Gamma distribution, with shape parameters and scale parameters respectively. and λ includes all variational parameters and Step 3.4.2: Use the maximum evidence lower bound to find the parameter λ so that the variational distribution Close to the true posterior distribution 6. The memory-aware graph data federated forgetting learning method according to claim 5 is characterized in that: In step 4, the model pruning and model performance recovery method, the specific steps are as follows: Step 4.1: The central server calculates the parameter weight Ω based on the forgotten user U' Prune the global GNN model W, and the pruning process is expressed as in represents the Hadamard product; Step 4.2: The central server will prune the model W * Send to other users of the same category as the forgotten user, and other users of the same category use the local graph interaction data to prune the model W * Fine-tune and upload the fine-tuned model to the central server; Step 4.3: The central server aggregates the fine-tuned models to obtain the final forget model.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor; characterized in that: When the processor executes the computer program, the electronic device executes the memory-aware graph data federated forgetting learning method as described in any one of claims 1 to 6.
8. A storage medium, comprising a computer program, which, when executed on an electronic device, enables the electronic device to execute the memory-aware graph data federated forgetting learning method as described in any one of claims 1 to 6.