AdaFedRec: Adaptive Heterogeneous Federated Recommender Systems Across Multi-Device Users

Although Federated Learning (FL), as a distributed machine learning paradigm, is promising for privacy-preserving model training in recommender systems, it suffers from low training performance due to heterogeneous hardware resources across user devices, especially among users with multiple devices. Specifically, i) the limited hardware capabilities of some user devices prevent them from participating in training large, highperformance models, while the limited number of users with high-performance devices results in insufficient data for training the large model. ii) Recommender systems are often deployed on multiple heterogeneous devices (e.g., personal computers and mobile phones) of each user, which makes it challenging to migrate personalized features of users for heterogeneous recommender models training across various uncertain devices. To address this issue, this paper presents a novel heterogeneous FedRec framework, AdaFedRec, that adaptively assigns heterogeneous models for local training based on the available resources of user devices and employs an adaptive heterogeneous model migration mechanism across multiple devices of a specific user. Specifically, AdaFedRec adopts an encoder-decoder-based heterogeneous model generation strategy for collaborative model training among heterogeneous devices. Additionally, AdaFedRec incorporates a knowledge-diffusion-based model-migration mechanism for training models across users' heterogeneous devices. The experimental results demonstrate that AdaFedRec achieves the best results across all datasets. Our code can be accessed at https://github.com/Iridescentttttt/AdaFedRec.