Fine-grained graph neural network recommendation method based on semi-decentralized federated learning

Abstract

The application discloses a fine-grained graph neural network recommendation method based on semi-decentralized federated learning, which constructs a semi-decentralized federated recommendation framework combining device-to-server and device-to-device under the premise that user original interaction data is always retained in a local device; a central server groups user and item embeddings uploaded by a client, allocates a predicted item node for each group, connects local self graphs isolated from each other of each client through the predicted item node, constructs a high-order local collaborative subgraph, and realizes high-order collaborative information transmission across clients; and an FGP module is used for adaptively aggregating global and local model parameters at an element level, so as to generate personalized models adapted to local preferences for different clients. The application effectively reduces the communication pressure of the central server, improves the system expansibility and recommendation accuracy, alleviates the data heterogeneity problem, and balances privacy protection and recommendation effect.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence recommendation technology, specifically to a fine-grained graph neural network recommendation method based on semi-decentralized federated learning. Background Technology

[0002] Recommender systems are widely used in e-commerce, video, news, and social media to predict user preferences based on historical user interactions. In recent years, graph neural networks have been applied to recommender systems, capturing high-order collaborative information by modeling user-item interaction graphs, thereby improving recommendation performance.

[0003] However, traditional graph neural network recommendation methods typically require centralized storage of user interaction data on servers, which can easily lead to privacy risks. Federated recommendation systems can alleviate privacy issues to some extent by keeping user data on local devices. However, existing federated recommendation methods still suffer from insufficient scalability, high communication overhead, and limited local device resources. Furthermore, because each user's local interaction graph is isolated from each other, it is difficult to fully utilize high-order collaborative information across users.

[0004] Furthermore, the data distribution and interests of different users vary significantly, making it difficult for a single global model to adapt to all clients. Existing personalized federated recommendation methods are typically coarse-grained, making it difficult to adaptively fuse global and local model information based on different users' local goals, thus limiting the effectiveness of personalized recommendations.

[0005] Therefore, those skilled in the art urgently need to develop a federated recommendation method that can improve system scalability, utilize high-order collaborative information, and achieve fine-grained personalized recommendations while protecting user privacy. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the prior art, the present invention at least solves the following technical problems: 1. Existing federated recommendation systems mostly rely on central servers for model aggregation and information exchange. When the number of clients is large, the communication overhead and server pressure are high. Some methods rely on third-party servers or complex encryption mechanisms, which are complex and pose a risk of privacy leakage. 2. In existing federated graph recommendation methods, the local self-graphs of each client are isolated from each other, making it difficult to utilize cross-user high-level collaborative information, while directly sharing the interaction graph will cause privacy leakage issues; 3. In federated recommendation scenarios, data heterogeneity is strong, the adaptability of a single global model is poor, and existing personalized methods are too coarse-grained and cannot accurately integrate effective global and local information.

[0007] To achieve the above objectives, this invention discloses a fine-grained graph neural network recommendation method based on semi-decentralized federated learning, which includes the following steps: S1: Each client constructs a local self-graph based on local user interaction data; S2: Each client uploads its local user embedding and local item embedding to the central server. The central server groups users and items according to the received embeddings, assigns a predicted item node to each group, and connects the local self-graphs of clients in the same group through the predicted item nodes to construct a higher-order local collaborative subgraph. S3: Clients within the same group communicate between devices through the predicted item node; S4: The central server sends global model parameters to each client. Each client uses the fine-grained personalization module to perform element-level adaptive aggregation of the global model parameters and local model parameters to generate a personalized local model. S5: Each client completes local training based on the personalized local model and generates recommendation results based on the embeddedness obtained from the training.

[0008] Furthermore, in step S1, the local self-graph includes user nodes, item nodes, and interaction edges between users and items; each client uses a local graph neural network model to propagate messages through the local self-graph to obtain the local user embedding and the local item embedding.

[0009] Furthermore, in step S2, the central server groups users and items based on the similarity of the received local user embeddings and local item embeddings, selects or generates corresponding predicted item nodes for each group, and establishes connections between the predicted item nodes and all user nodes within the group.

[0010] Furthermore, in step S3, the clients do not share the original interaction records and local self-graphs, but only exchange the aggregated node embedding information, and use the message propagation of the higher-order local collaborative subgraph to enhance the expressive power of the local graph neural network model.

[0011] Furthermore, in step S4, the element-level adaptive aggregation process of the fine-grained personalization module is executed according to the following formula:

[0012] in, For the client The element-level adaptive aggregated weight matrix corresponding to the high-level model. For the clipping operation operator, For the client No. The local model parameters of the wheel, To form a weight matrix with the same shape as the low-level parameters of the model and all elements being 1. For the first Global model parameters of the wheel, For Hadamard products; the client adaptively aggregates weights for higher layers based on a% of local sampled data. The training process is performed, and its values ​​are kept within a preset range by pruning constraints.

[0013] Furthermore, the fine-grained personalization module employs different aggregation methods at the low and high levels of the model: for the low level, it utilizes... The client primarily inherits global model parameters to retain the common representation capabilities shared across clients; for higher levels of the model, element-level adaptive aggregation weights are learned based on the client's local data. It performs personalized weighting on the differences between global model parameters and local model parameters, thereby enabling the aggregated local model to better adapt to the local user preferences of the client.

[0014] Furthermore, in step S2, before uploading the local item embedding to the central server, the client performs a trimming process on the local item embedding and adds noise perturbation.

[0015] Furthermore, in step S5, after each client completes local training based on the personalized local model, it updates the local user embedding and the local item embedding. The specific method for generating recommendation results is as follows: based on the updated local user embedding and local item embedding, the user's preference score for candidate items is calculated, and recommendation results are generated by sorting them according to the scores.

[0016] Furthermore, the central server maintains a global item embedding table. After receiving the embedding uploaded by the client, it is only used to update the global item embedding table and build user-item grouping relationships, and does not store the user's original interaction data.

[0017] Furthermore, during each training round, after each client completes local training, it uploads the updated local user embeddings and local item embeddings to the central server to enter the next training iteration.

[0018] This invention achieves at least the following beneficial technical effects: 1. A semi-decentralized federated recommendation framework, pFedSG, is proposed, which combines device-to-server and device-to-device collaboration, reducing the dependence on a central server and improving the system scalability in large-scale client scenarios. At the same time, it avoids directly uploading users' original interaction data, enhancing privacy protection capabilities. Moreover, the framework can be adapted to various federated recommendation models based on graph neural networks. 2. An isolated self-graph connection mechanism based on predicted item nodes is proposed. Without sharing the original data, the model can utilize cross-user high-order collaborative information, which alleviates the problem of insufficient local graph information, improves recommendation accuracy and model expressive power, and maintains the privacy protection characteristics in the federated recommendation scenario. 3. A fine-grained personalization module is proposed, which effectively integrates global common information and local personalized information through element-level adaptive aggregation, thereby improving the model's adaptability to the personalized preferences of different clients and alleviating the data heterogeneity problem in federated learning.

[0019] This invention does not rely on centralized storage of complete user interaction graphs, nor does it require clients to directly share raw user behavior data. It is suitable for deployment on mobile terminals, edge devices, and large-scale distributed user environments, and can be applied to various online recommendation scenarios such as e-commerce recommendations, video recommendations, news recommendations, social media recommendations, and book recommendations. This invention achieves recommendation model training and personalized recommendations through a hybrid collaborative approach without requiring raw user interaction data to leave the local device. It has good privacy protection capabilities and practical application value, while achieving a good balance between privacy protection, system scalability, and recommendation accuracy, making it highly practical for engineering and widespread application. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the fine-grained graph neural network recommendation method based on semi-decentralized federated learning of the present invention. Figure 2 This is a schematic diagram of the semi-decentralized federated recommendation collaboration mode of the present invention; Figure 3 This is a schematic diagram of the parameter aggregation process of the fine-grained personalization module of the present invention. Detailed Implementation

[0021] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0022] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0023] This invention proposes a fine-grained graph neural network recommendation method based on semi-decentralized federated learning, such as... Figure 1As shown, the overall process includes five core steps: local self-graph construction, grouping and higher-order subgraph construction, inter-device message passing, fine-grained model aggregation, local training, and recommendation generation. S1: Each client constructs a local self-graph based on local user interaction data; S2: Each client uploads its local user embedding and local item embedding to the central server. The central server groups users and items according to the received embeddings, assigns a predicted item node to each group, and connects the local self-graphs of clients in the same group through the predicted item nodes to construct a higher-order local collaborative subgraph. S3: Clients within the same group communicate between devices through the predicted item node; S4: The central server sends global model parameters to each client. Each client uses the Fine-grained Personalization (FGP) module to perform element-level adaptive aggregation of the global model parameters and local model parameters to generate a personalized local model. S5: Each client completes local training based on the personalized local model and generates recommendation results based on the embeddedness obtained from the training.

[0024] like Figure 2 As shown, this method employs a semi-decentralized collaboration model combining device-to-server and device-to-device approaches: the central server maintains a global item embedding table, handling only user and item grouping, predicted item node allocation, and global model distribution, without storing original user interaction data; multiple clients are divided into different groups based on embedding similarity, and clients within the same group establish indirect connections through predicted item nodes, completing message interaction only at the embedding level. The detailed execution process of this collaboration model is as follows: Each client constructs a local self-graph based on local user interaction data. This local self-graph includes user nodes, item nodes, and interaction edges between users and items. Original user behavior data, such as clicks, browsing, ratings, and purchases, is always stored locally on the device and is not uploaded to the server or other clients. Each client uses a local graph neural network model to perform message propagation on the local self-graph, obtaining local user embeddings and local item embeddings.

[0025] Each client uploads its calculated local user embeddings and local item embeddings to the central server. The central server maintains a global item embedding table. After receiving the embeddings uploaded by the clients, it only updates the global item embedding table and constructs user-item grouping relationships; it does not store the original user interaction data. The central server groups users and items based on the similarity of the received local user embeddings and local item embeddings. For each group, it selects or generates a corresponding predicted item node. The predicted item node establishes a connection with all user nodes within the group. Through the predicted item node, the originally isolated local self-graphs of clients within the same group are connected, constructing a higher-order local collaborative subgraph.

[0026] As an optional implementation, to further reduce the risk of privacy leakage, the client can trim the local item embedding and add noise perturbation before uploading the local item embedding to the central server, so as to reduce the possibility of deducing user privacy data from the embedded information; this implementation is particularly suitable for recommendation scenarios with high privacy protection requirements.

[0027] Clients within the same group communicate via the predicted item nodes for device-to-device messaging. Clients do not share original interaction records or local self-graphs; they only exchange aggregated node embedding information. In this way, clients can leverage higher-order collaboration information from similar users while protecting privacy, thereby enhancing the expressive power of the local graph neural network model.

[0028] During each training round, the central server distributes global model parameters to each client. Each client then uses the FGP module to perform element-level adaptive aggregation of the global model parameters and local model parameters to generate a personalized local model.

[0029] like Figure 3 As shown, the parameter aggregation process of the FGP module is as follows: Based on the previous round of local model parameters, the client calculates the difference between the global model parameters and the local model parameters, and then performs a weighted fusion of these differences using element-level aggregation weights to obtain personalized model parameters adapted to the current client. The specific calculation formula is as follows:

[0030] in, For the client The element-level adaptive aggregated weight matrix corresponding to the high-level model. For the clipping operation operator, For the client No. The local model parameters of the wheel, To form a weight matrix with the same shape as the low-level parameters of the model and all elements being 1. For the first Global model parameters of the wheel, For Hadamard products; the client adaptively aggregates weights for higher layers based on a% of local sampled data. The training process is performed, and its values ​​are kept within a preset range by pruning constraints.

[0031] The fine-grained personalization module employs different aggregation methods at the low and high levels of the model: for the low level, it utilizes... The client primarily inherits global model parameters, retaining general representation capabilities; for higher levels of the model, element-level adaptive aggregation weights are learned based on the client's local data. It personalizes the weighting of the differences between global model parameters and local model parameters to adapt to local user preferences, thereby utilizing both global common information and local personalized information.

[0032] After each client completes local training based on the personalized local model, it updates the local user embedding and local item embedding. Based on the updated local user embedding and local item embedding, it calculates the user's preference score for the candidate items and generates recommendation results by sorting them according to the scores.

[0033] During each training round, after each client completes local training, it uploads the updated local user embeddings and local item embeddings to the central server to enter the next training iteration.

[0034] This method can be applied to scenarios such as e-commerce recommendations, video recommendations, news recommendations, social media recommendations, music recommendations, and book recommendations. Taking e-commerce recommendations as an example, each user terminal stores local user behavior data such as browsing, favorites, clicks, and purchases, and constructs a local self-graph between users and items (products). This method can group similar users and connect similar users using predicted item nodes, enabling the local model to learn richer collaborative purchasing and preference information. In the recommendation stage, the client calculates the user's preference score for candidate items based on the personalized model and generates an item recommendation list.

[0035] In experimental verification, this invention was applied to four real-world recommendation datasets and compared with existing federated recommendation methods and graph neural network recommendation methods. Experimental results show that this invention can effectively improve recommendation performance. Compared with the best baseline method, this invention improves the HR metric by approximately 7.37% and the NDCG metric by approximately 6.51%, demonstrating good recommendation performance and practical value.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0037] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A fine-grained graph neural network recommendation method based on semi-decentralized federated learning, characterized in that, Includes the following steps: S1: Each client constructs a local self-graph based on local user interaction data; S2: Each client uploads its local user embedding and local item embedding to the central server. The central server groups users and items according to the received embeddings, assigns a predicted item node to each group, and connects the local self-graphs of clients in the same group through the predicted item nodes to construct a higher-order local collaborative subgraph. S3: Clients within the same group communicate between devices through the predicted item node; S4: The central server sends global model parameters to each client. Each client uses the fine-grained personalization module to perform element-level adaptive aggregation of the global model parameters and local model parameters to generate a personalized local model. S5: Each client completes local training based on the personalized local model and generates recommendation results based on the embeddedness obtained from the training.

2. The fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, in step S1, the local self-graph includes user nodes, item nodes, and interaction edges between users and items; each client uses a local graph neural network model to perform message propagation on the local self-graph to obtain the local user embedding and the local item embedding.

3. In the fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, in step S2, the central server groups users and items according to the similarity of the received local user embeddings and local item embeddings, selects or generates corresponding predicted item nodes for each group, and establishes connections between the predicted item nodes and all user nodes in the group.

4. In the fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 2, in step S3, the clients do not share the original interaction records and local self-graphs, but only exchange the aggregated node embedding information, and use the message propagation of the higher-order local collaborative subgraph to enhance the expressive power of the local graph neural network model.

5. The fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, in step S4, the element-level adaptive aggregation process of the fine-grained personalization module is executed according to the following formula: in, For the client The element-level adaptive aggregated weight matrix corresponding to the high-level model. For the clipping operation operator, For the client No. The local model parameters of the wheel, To form a weight matrix with the same shape as the low-level parameters of the model and all elements being 1. For the first Global model parameters of the wheel, For Hadamah accumulation; The client adaptively aggregates weights for higher layers based on a% of local sampled data. The training process is performed, and its values ​​are kept within a preset range by pruning constraints.

6. The fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 5, wherein the fine-grained personalization module adopts different aggregation methods in the low-level and high-level layers of the model: the low-level layer utilizes... The client inherits global model parameters; the higher-level model learns element-level adaptive aggregation weights based on the client's local data. It performs personalized weighting on the differences between global model parameters and local model parameters.

7. In the fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, in step S2, before uploading the local item embedding to the central server, the client performs pruning processing on the local item embedding and adds noise perturbation.

8. In the fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, in step S5, after each client completes local training based on the personalized local model, it updates the local user embedding and local item embedding; the specific way to generate recommendation results is as follows: based on the updated local user embedding and local item embedding, calculate the user's preference score for candidate items, and generate recommendation results according to the scores.

9. The fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, wherein the central server maintains a global item embedding table, and after receiving the embedding uploaded by the client, it is only used to update the global item embedding table and construct user-item grouping relationships, and does not store the original user interaction data.

10. The fine-grained graph neural network recommendation method based on semi-decentralized federated learning according to claim 1, wherein in each training round, after each client completes local training, it uploads the updated local user embedding and local item embedding to the central server to enter the next training iteration.

Images