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

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

TL;DR

FCUCR is a new federated learning framework that tackles the "forgetting" problem in recommender systems. By combining a time-aware self-distillation strategy with a privacy-preserving mechanism to share "intent prototypes" between users, it allows models to adapt to evolving user behaviors without losing sight of long-term preferences or compromising data privacy.

What Problem Does This Paper Address?

Most modern recommender systems are cloud-based, requiring massive amounts of user data to be centralized for training. Federated Learning (FL) offers a privacy-preserving alternative by keeping data on-device, but it faces two major hurdles in the real world:

  1. Temporal Forgetting: Users aren't static. Our interests shift over time. Most FL models look at fixed snapshots of data; when they update incrementally on new interactions, they often suffer from "catastrophic forgetting," losing the signal of a user’s long-term, foundational preferences.
  2. Weakened Collaborative Personalization: Recommendation is inherently social—it works best when the system learns that "users like you also liked this." However, in a federated setting where data is heterogeneous and private, it is incredibly difficult to find these cross-user patterns without leaking sensitive information.

The authors introduce FCUCR (Federated Continual User-Centric Recommendation) to bridge the gap between privacy, continual adaptation, and collaborative intelligence.

Methods

The FCUCR framework introduces two primary innovations:

  • Time-Aware Self-Distillation: To prevent the model from forgetting the past as it learns from new sessions, the authors implement a self-distillation loss. At each time step, the current representation module is regularized to stay semantically consistent with its state from the previous time step. This acts as an "implicit" memory, allowing the model to inherit historical knowledge without the need to store old, bulky interaction data on the device.
  • Inter-User Prototype Transfer: To restore the "collaborative" power of recommendation, the server maintains a dynamic Knowledge Base (KB) of "prototypes"—compact, numerical representations of user intent. Instead of sharing raw logs, clients upload their local prototypes. The server then identifies and sends back the top-$k$ most similar prototypes from other users. This allows a client's model to "borrow" insights from similar users to enrich its local representations while keeping the final decision logic (the prediction module) strictly private and local.

The architecture is modular, decoupling representation learning (which captures item semantics and is shared/federated) from the prediction module (which captures individual decision logic and stays private).

Datasets

The authors evaluated FCUCR on four diverse, real-world benchmark datasets, preprocessed into temporal sessions to simulate evolving preferences:

  • XING: Anonymized user-job interactions.
  • RetailRocket: Six months of e-commerce browsing logs.
  • LastFM: Detailed music listening histories.
  • Tmall: Large-scale user shopping interactions.

Key Insights and Ideas

  • Semantic Stability vs. Interest Shifts: While a user's specific interest might shift from "running shoes" to "hiking boots," their underlying cognitive interpretation of what makes a "good product" often remains stable. Self-distillation preserves this stable semantic "vibe."
  • Prototypes as Privacy-Preserving Bridges: You don't need to see a user's entire history to help them. By sharing abstract "intent prototypes," the system can find "behavioral twins" across the network.
  • Decoupling is Key: By keeping the prediction head local and only federating the representation layer, the model can benefit from global knowledge while maintaining highly personalized, user-specific decision logic.

Key Results

  • Superior Performance: FCUCR consistently outperformed state-of-the-art federated (e.g., FedRAP, GPFedRec) and session-based (e.g., Fed-DMI-GNN) baselines across all four datasets. On the XING dataset, it improved the Hit Rate (HR@10) significantly compared to standard federated sequential models.
  • Resistance to Forgetting: Visualization of performance over time (Figure 2a) showed that while baseline models' accuracy on early sessions decayed as they learned new data, FCUCR maintained high accuracy, proving its "continual" strength.
  • Robust Privacy: The authors tested the framework with Local Differential Privacy (LDP). Even when injecting significant noise to enhance privacy, FCUCR maintained over 95% of its original accuracy, demonstrating high practical viability for privacy-sensitive applications.
  • Efficiency: The framework allows for "sliding window" knowledge bases and reduced transfer frequencies, which significantly lowers the computational and communication overhead without tanking performance.

Limitations and Future Work

While the paper demonstrates strong results, a few areas remain for exploration:

  • Prototype Fusion: Currently, the fusion of retrieved prototypes is relatively simple (averaging or basic dimensionality reduction). More complex attention-based fusion could potentially capture even subtler cross-user relationships.
  • Extreme Heterogeneity: While the model handles data diversity well, extremely "outlier" users might find fewer similar prototypes in the KB, potentially leading to a "cold-start" issue for very unique behavioral patterns.
  • Scaling the Knowledge Base: As the number of users grows into the millions, managing the server-side Knowledge Base will require more advanced indexing and pruning strategies to maintain the sub-second retrieval times required for real-time recommendation.

Summary updated

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.
BibTeX
@misc{zhang2026learning,
  author = {Chunxu Zhang and Zhiheng Xue and Guodong Long and Weipeng Zhang and Bo Yang},
  title = {{Learning Evolving Preferences: A Federated Continual Framework for User-Centric Recommendation}},
  year = {2026},
  eprint = {2603.17315},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.17315},
}
Figure