Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Evolving Preferences: A Federated Continual Framework for User-Centric Recommendation

Chunxu Zhang, Zhiheng Xue, Guodong Long, Weipeng Zhang, Bo Yang

Abstract

User-centric recommendation has become essential for delivering personalized services, as it enables systems to adapt to users' evolving behaviors while respecting their long-term preferences and privacy constraints. Although federated learning offers a promising alternative to centralized training, existing approaches largely overlook user behavior dynamics, leading to temporal forgetting and weakened collaborative personalization. In this work, we propose FCUCR, a federated continual recommendation framework designed to support long-term personalization in a privacy-preserving manner. To address temporal forgetting, we introduce a time-aware self-distillation strategy that implicitly retains historical preferences during local model updates. To tackle collaborative personalization under heterogeneous user data, we design an inter-user prototype transfer mechanism that enriches each client's representation using knowledge from similar users while preserving individual decision logic. Extensive experiments on four public benchmarks demonstrate the superior effectiveness of our approach, along with strong compatibility and practical applicability. Code is available.

Learning Evolving Preferences: A Federated Continual Framework for User-Centric Recommendation

Abstract

User-centric recommendation has become essential for delivering personalized services, as it enables systems to adapt to users' evolving behaviors while respecting their long-term preferences and privacy constraints. Although federated learning offers a promising alternative to centralized training, existing approaches largely overlook user behavior dynamics, leading to temporal forgetting and weakened collaborative personalization. In this work, we propose FCUCR, a federated continual recommendation framework designed to support long-term personalization in a privacy-preserving manner. To address temporal forgetting, we introduce a time-aware self-distillation strategy that implicitly retains historical preferences during local model updates. To tackle collaborative personalization under heterogeneous user data, we design an inter-user prototype transfer mechanism that enriches each client's representation using knowledge from similar users while preserving individual decision logic. Extensive experiments on four public benchmarks demonstrate the superior effectiveness of our approach, along with strong compatibility and practical applicability. Code is available.
Paper Structure (25 sections, 8 equations, 10 figures, 5 tables, 1 algorithm)

This paper contains 25 sections, 8 equations, 10 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Continual Recommendation
  4. Federated Recommender System
  5. Preliminary
  6. Methodology
  7. Time-Aware Self-Distillation Strategy
  8. Inter-User Prototype Transfer Mechanism
  9. Optimization Objective and Algorithm
  10. Discussion on Practical Adaptability
  11. Experiment
  12. Experimental Setup
  13. Overall Comparison
  14. Component Effectiveness Analysis
  15. Hyper-Parameter Analysis
  16. ...and 10 more sections

Figures (10)

  • Figure 1: Overall framework of our proposed FCUCR. We focus on continual recommendation, where clients receive new interaction sessions over time. To mitigate temporal forgetting, we apply a time-aware self-distillation strategy that uses the representation module from the previous time step to guide the current session’s learning. Additionally, We introduce an inter-user prototype transfer mechanism where the server maintains a dynamic prototype knowledge base and retrieves the top-$k$ similar prototypes per client to enable collaborative personalization. The figure illustrates the procedure at time step $t$.
  • Figure 2: Component effectiveness analysis on XING dataset.
  • Figure 3: Effect of two hyper-parameters on model performance (HR@10) on the XING dataset.
  • Figure 4: Effect of two hyper-parameters on model performance (HR@10) on the RetailRocket dataset.
  • Figure 5: Effect of two hyper-parameters on model performance (HR@10) on the LastFM dataset.
  • ...and 5 more figures