GPFedRec: Graph-guided Personalization for Federated Recommendation

TL;DR

GPFedRec addresses privacy-preserving federated recommendation by constructing a server-side user relation graph from locally updated item embeddings and applying graph-guided aggregation to learn user-specific item embeddings without accessing private interactions. The server derives $r_i$ via a 1‑layer Graph Convolution Network on an adaptive adjacency $\mathcal{A}$, computes $q_{global}$ as a degree-weighted aggregation of inputs, and clients update with $\mathcal{L}_{total} = \mathcal{L}_i(\theta_i; \mathcal{Y}_{um}, \hat{\mathcal{Y}}_{um}) + \lambda \mathcal{R}(q_i, r_i)$. Training alternates server and client updates, and local differential privacy can be applied by adding Laplacian noise to $q_i$. Experiments on five datasets show state-of-the-art performance and broad compatibility with other FedRec models, validating the practical impact of graph-guided personalization in privacy-preserving federated settings.

Abstract

The federated recommendation system is an emerging AI service architecture that provides recommendation services in a privacy-preserving manner. Using user-relation graphs to enhance federated recommendations is a promising topic. However, it is still an open challenge to construct the user-relation graph while preserving data locality-based privacy protection in federated settings. Inspired by a simple motivation, similar users share a similar vision (embeddings) to the same item set, this paper proposes a novel Graph-guided Personalization for Federated Recommendation (GPFedRec). The proposed method constructs a user-relation graph from user-specific personalized item embeddings at the server without accessing the users' interaction records. The personalized item embedding is locally fine-tuned on each device, and then a user-relation graph will be constructed by measuring the similarity among client-specific item embeddings. Without accessing users' historical interactions, we embody the data locality-based privacy protection of vanilla federated learning. Furthermore, a graph-guided aggregation mechanism is designed to leverage the user-relation graph and federated optimization framework simultaneously. Extensive experiments on five benchmark datasets demonstrate GPFedRec's superior performance. The in-depth study validates that GPFedRec can generally improve existing federated recommendation methods as a plugin while keeping user privacy safe. Code is available to ease reproducibility

Figures (6)

• Figure 1: Performance comparison of diverse user relationship graphs-enhanced federated recommendation model on the MovieLens-100K dataset. Backbone denotes the current state-of-the-art federated recommendation model.
• Figure 2: The framework of GPFedRec. There are four steps in each communication round: ① For the local recommendation model trained on each client, it initializes the item embedding with the globally shared item embedding. Then it takes the user-specific item embedding as a regularizer $\mathcal{R}(q,r_i)$ together with the loss of the recommendation task $\mathcal{L}_i(\mathcal{Y},\hat{\mathcal{Y}})$ as the optimization objective $\mathcal{L}_{total}$. ② The client uploads the locally updated item embedding $q$ to the server. ③ For the server, it first constructs a user relationship graph based on the received item embeddings. Then, it performs the graph-guided aggregation to achieve user-specific item embeddings $\{r_i\}_{i=1}^n$ and meanwhile calculates the globally shared item embedding depicting the popular preference. ④ The server distributes both the globally shared and user-specific item embeddings to the clients for the next round of optimization.
• Figure 3: Effect of the threshold of neighborhood selection. We show the results of both metrics on MovieLens-100K.
• Figure 4: Effect of coefficient of the regularization term. We show the results of both metrics on MovieLens-100K.
• Figure 5: Effect of embedding size. We show the results of both metrics on MovieLens-100K.