Abstract:Graph neural network (GNN) has emerged as a state-of-the-art solution for item recommendation. However, existing GNN-based recommendation methods rely on a centralized storage of fragmented user-item interaction sub-graphs and training on an aggregated global graph, which will lead to privacy concerns. As a response, some recent works develop GNN-based federated recommendation methods by exploiting decentralized and fragmented user-item sub-graphs in order to preserve user privacy. However, due to privacy constraints, the graph convolution process in existing federated recommendation methods is incomplete compared with the centralized counterpart, causing a degradation of the recommendation performance. In this paper, we propose a novel lossless and privacy-preserving graph convolution network (LP-GCN), which fully completes the graph convolution process with decentralized user-item interaction sub-graphs while ensuring privacy. It is worth mentioning that its performance is equivalent to that of the non-federated (i.e., centralized) counterpart. Moreover, we validate its effectiveness through both theoretical analysis and empirical studies. Extensive experiments on three real-world datasets show that our LP-GCN outperforms the existing federated recommendation methods. The code will be publicly available once the paper is accepted.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: in the federated learning framework, how to achieve no loss in the performance of the graph neural network (GNN) recommendation system while protecting user privacy. Specifically, the existing federated recommendation methods based on GNN lead to an incomplete graph convolution process due to privacy constraints, thus affecting the recommendation performance. To this end, the authors propose a lossless and privacy - protected graph convolution network (LP - GCN) to ensure the complete execution of the graph convolution process on the decentralized user - item interaction sub - graph while guaranteeing privacy.
### Problem Background
With the popularization of the Internet and mobile devices, it has become increasingly difficult to find information that users are interested in from massive online data in fields such as e - commerce and fintech. Recommendation systems help alleviate this problem by filtering out irrelevant information and recommending items that match users' interests. Traditional recommendation models rely on servers to collect and manage users' historical data such as browsing, clicking, and purchasing to train centralized models. However, these behavioral data usually contain users' private information, and users may be unwilling to share this information with the server.
To solve this problem, federated learning, as a distributed machine - learning framework, provides an effective solution. Federated learning allows multiple participants to collaborate in training models, while the original data remains local and will not be obtained by other participants. Therefore, federated learning can effectively protect user privacy and, in many cases, can achieve performance comparable to that of centralized models.
### Existing Challenges
Although previous studies have proposed federated recommendation frameworks based on GNN, they face two main problems:
1. **Risk of Privacy Leakage**: Existing methods expand users' sub - graphs by encrypting item IDs or using inter - group hiding to obtain high - order neighborhood information. However, encrypting item IDs relies on third - party servers, and inter - group hiding is vulnerable to inference attacks.
2. **Performance Degradation**: Due to privacy limitations, the graph convolution process is incomplete, resulting in performance inferior to that of centralized models.
### Solution
To address the above challenges, the authors propose a novel lossless federated graph recommendation framework named LP - GCN. The main contributions of LP - GCN include:
1. **Complete Execution of the Graph Convolution Process**: LP - GCN can completely execute the graph convolution process, including forward propagation and backward propagation, on the decentralized user - item interaction sub - graph, thereby achieving performance comparable to that of centralized models.
2. **Combination of Hybrid Encryption of Item IDs and Virtual Item Obfuscation**: Through this method, the server can construct a global user - item graph without leaking user privacy and assist participating clients in graph expansion. Even if the server colludes with some users, this method can still effectively protect user privacy.
3. **Embedding Synchronization Mechanism**: By calculating the embeddings of shared nodes in specific clients and synchronizing them through the server to other clients containing the same nodes, it is ensured that no information is lost during the graph convolution process.
4. **Theoretical Proof and Experimental Verification**: The authors provide theoretical analysis and experimental proof, indicating that after federalizing the existing algorithms, LP - GCN can achieve performance comparable to that of centralized models.
### Experimental Results
The authors conducted extensive experiments on three real - world datasets to evaluate the recommendation performance and communication costs. The experimental results show that LP - GCN not only outperforms existing methods in performance but also performs excellently in protecting user privacy.
In conclusion, this paper aims to solve the privacy and performance problems in the existing federated recommendation methods based on GNN by proposing LP - GCN, achieving a lossless and privacy - protected recommendation system.