Lossless and Privacy-Preserving Graph Convolution Network for Federated Item Recommendation

TL;DR

This work tackles privacy concerns in graph-based federated item recommendation by proposing LP-GCN, a lossless federated graph convolution network that completes both forward and backward propagation without leaking sensitive data. It achieves this via a hybrid encryption scheme and an embedding-synchronization mechanism that allows construction and propagation over a global, encrypted item graph while preserving privacy—even under partial server collusion. Theoretical analysis and experiments on Gowalla, Yelp2018, and Amazon-Book demonstrate that LP-GCN matches the performance of centralized counterparts and outperforms existing federated methods, with acceptable communication costs. The framework thus enables privacy-preserving, high-performance GNN-based recommendations in distributed settings, opening avenues for broader deployment and future work on auxiliary data and fairness.

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.

Figures (13)

• Figure 1: Illustration of the centralized and federated versions of GNN-based recommendation.
• Figure 2: Illustration of the centralized and federated versions of forward propagation and backward propagation w.r.t. item $i$.
• Figure 3: The overview of our LP-GCN.
• Figure 4: The overall framework of our LP-GCN, which consists of four states. In state 1, each client has the $0$-th user embeddings, and the convolution-items in convolution-clients have the $0$-th item embeddings, while ordinary items do not have any embeddings. After forward propagation, the model moves to state 2, where all clients have user and item embeddings of all layers. Subsequently, clients use these embeddings to construct local losses (i.e., the green lines) and compute gradients for all layers' embeddings (i.e., the red lines) and then transitions to state 3. Finally, through backward propagation, all the intermediate gradients would be propagated back to the learnable parameters, as shown in state 4.
• Figure 5: An example of splitting clients and items.