Federated Recommendation with Additive Personalization

TL;DR

FedRAP tackles privacy-preserving federated recommendation by introducing a bipartite personalization architecture that combines a sparse global item embedding $\mathbf{C}$ with user-specific local item embeddings $\mathbf{D}^{(i)}$, yielding personalized item representations $\mathbf{D}^{(i)}+\mathbf{C}$. It employs a curriculum that gradually shifts from full personalization to additive personalization by increasing regularization strengths and applying an $\ell_1$ constraint on $\mathbf{C}$ to reduce communication costs. The method is optimized via alternating updates across clients and server, with differential privacy applied to the gradients of $\mathbf{C}$, and is empirically validated on six real-world datasets where FedRAP outperforms other federated methods and approaches centralized upper bounds. The results demonstrate improved personalization under FL with controllable communication overhead and robust privacy protections, highlighting FedRAP’s practical relevance for privacy-conscious recommender systems.

Abstract

Building recommendation systems via federated learning (FL) is a new emerging challenge for advancing next-generation Internet service and privacy protection. Existing approaches train shared item embedding by FL while keeping the user embedding private on client side. However, item embedding identical for all clients cannot capture users' individual differences on perceiving the same item and thus leads to poor personalization. Moreover, dense item embedding in FL results in expensive communication cost and latency. To address these challenges, we propose Federated Recommendation with Additive Personalization (FedRAP), which learns a global view of items via FL and a personalized view locally on each user. FedRAP enforces sparsity of the global view to save FL's communication cost and encourages difference between the two views through regularization. We propose an effective curriculum to learn the local and global views progressively with increasing regularization weights. To produce recommendations for an user, FedRAP adds the two views together to obtain a personalized item embedding. FedRAP achieves the best performance in FL setting on multiple benchmarks. It outperforms recent federated recommendation methods and several ablation study baselines.

Figures (8)

• Figure 1: The framework of FedRAP. For each user-$i$, FedRAP utilizes its ratings $r_i$ as labels to locally train a user embedding $\mathbf{u}_i$ and a local item embedding $\mathbf{D}^{(i)}$. By adding $\mathbf{D}^{(i)}$ to the global item embedding $\mathbf{C}$ from the server, i.e., additive personalization, FedRAP creates a personalized item embedding based on both shared knowledge and personal perspective and thus produces better recommendations. Since clients and server only communicate $\mathbf{C}$, FedRAP enforce its sparsity to reduce the communication cost and encourage its difference to $\mathbf{D}^{(i)}$ so they are complementary.
• Figure 2: Evaluation metrics (%) on six real-world recommendation datasets. CentRAP is a centralized version (upper bound) of FedRAP. FedRAP outperforms all the FL methods by a large margin.
• Figure 3: Convergence and efficiency comparison of four methods on the ML-100K dataset.
• Figure 4: Different curriculum for $\lambda_(a,v_1)$ in Eq. (\ref{['eq:obj']}) with $v_1=0.1$ and $a$ to be the iteration.
• Figure 5: Ablation study investigating the effectiveness of FedRAP on the ML-100K dataset.