Towards Personalized Federated Multi-Scenario Multi-Task Recommendation

TL;DR

This paper introduces PF-MSMTrec, a personalized federated multi-scenario multi-task recommendation framework that effectively handles multiple tasks and scenarios while addressing data privacy and optimization conflicts.

Contribution

The paper proposes a novel federated learning framework with scenario-specific parameters, conflict coordination, and personalized aggregation for multi-task recommendation.

Findings

Outperforms state-of-the-art methods on public datasets.

Effectively handles multi-scenario multi-task recommendation.

Maintains data privacy through federated learning.

Abstract

In modern recommender systems, especially in e-commerce, predicting multiple targets such as click-through rate (CTR) and post-view conversion rate (CTCVR) is common. Multi-task recommender systems are increasingly popular in both research and practice, as they leverage shared knowledge across diverse business scenarios to enhance performance. However, emerging real-world scenarios and data privacy concerns complicate the development of a unified multi-task recommendation model. In this paper, we propose PF-MSMTrec, a novel framework for personalized federated multi-scenario multi-task recommendation. In this framework, each scenario is assigned to a dedicated client utilizing the Multi-gate Mixture-of-Experts (MMoE) structure. To address the unique challenges of multiple optimization conflicts, we introduce a bottom-up joint learning mechanism. First, we design a parameter template to decouple the expert network parameters, distinguishing scenario-specific parameters as shared knowledge for federated parameter aggregation. Second, we implement personalized federated learning for each expert network during a federated communication round, using three modules: federated batch normalization, conflict coordination, and personalized aggregation. Finally, we conduct an additional round of personalized federated parameter aggregation on the task tower network to obtain prediction results for multiple tasks. Extensive experiments on two public datasets demonstrate that our proposed method outperforms state-of-the-art approaches. The source code and datasets will be released as open-source for public access.