A Systematic Survey on Federated Sequential Recommendation

TL;DR

This survey addresses the privacy challenges of sequential recommendation by introducing Federated Sequential Recommendation (FedSR), which trains a global sequential model across users without sharing raw data. It provides a two-level taxonomy of solutions—Model-Level (Parameter Decomposition and LLM Foundation) and Device-Level (Communication Optimization and Aggregation Balance)—with concrete techniques and representative works for each. The authors identify core challenges such as data heterogeneity, communication costs, and privacy risks in updates, and outline future directions including personalized strategies, LLM integration under FL, multi-modal fusion, cold-start adaptation, and federated continual learning. The work aims to guide researchers and practitioners toward privacy-preserving, scalable, and effective FedSR systems with practical insights for real-world applications.

Abstract

Sequential recommendation is an advanced recommendation technique that utilizes the sequence of user behaviors to generate personalized suggestions by modeling the temporal dependencies and patterns in user preferences. However, it requires a server to centrally collect users' data, which poses a threat to the data privacy of different users. In recent years, federated learning has emerged as a distributed architecture that allows participants to train a global model while keeping their private data locally. This survey pioneers Federated Sequential Recommendation (FedSR), where each user joins as a participant in federated training to achieve a recommendation service that balances data privacy and model performance. We begin with an introduction to the background and unique challenges of FedSR. Then, we review existing solutions from two levels, each of which includes two specific techniques. Additionally, we discuss the critical challenges and future research directions in FedSR.

Figures (4)

• Figure 1: The framework of FedSR. In FedSR, each user participates in the federated training with the local model and dataset. Firstly, the user trains the local model with the private dataset locally and then uploads the model parameters to the server. The server aggregates the local models and broadcasts the global model to the users participating in the next communication round.
• Figure 2: Taxonomy of FedSR. We classify them into two subcategories and four primary techniques, i.e., model-level (parameter decomposition and LLM foundation) and device-level (communication optimization and aggregation balance). Different colors indicate categories, and we list representative works in the boxes.
• Figure 3: Model-level methods, including parameter decomposition and LLM foundation. Parameter decomposition captures universal patterns across clients while leveraging local parameters to adapt to local dataset, and LLM foundation guides knowledge retention and parameter updates by integrating knowledge from LLM.
• Figure 4: Device-level methods, including communication optimization and aggregation balance. Communication optimization focuses on reducing the communication cost between clients and the server, and the aggregation balance often strategically aggregates models to maximize the optimization of the global model.