Federated Graph Neural Network-based Recommender Systems

The School of EECS is hosting the following PhD Thesis Review Seminar:

Federated Graph Neural Network-based Recommender Systems

Speaker: Liang Qu 
Host: Prof Hongzhi Yin

Abstract: Recommender systems have become an integral part of our digital lives, providing personalized suggestions across various domains such as e-commerce and video streaming, by filtering massive datasets to match user preferences. Recent advancements have highlighted the efficacy of Graph Neural Networks (GNNs) in recommender systems, achieving state-of-the-art performance by naturally mapping user-item interactions into bipartite graphs and utilizing GNNs to learn sophisticated user/item embeddings. However, this approach raises significant privacy concerns due to the necessity of aggregating user interaction data. In light of these concerns, federated learning has emerged as a promising privacy-preserving paradigm that retains data on user devices while collaboratively training models via model parameter exchange. Despite its advancements, federated GNN-based recommender systems face unique challenges, including performance degradation, inflexible architectures, slow convergence, and the demand for substantial local storage resources.

This thesis introduces several innovations to address these challenges. Firstly, we propose a semi-decentralized federated ego graph learning for recommendation architecture to counteract performance loss due to the inability of GNNs to learn from high-order graph structures in local one-hop subgraphs. By generating fake common nodes to connect lower-order subgraphs, this method enables GNNs to learn higher-order information without sharing user data. Secondly, we introduce a personalized privacy federated GNN-based recommendation architecture that allows users to tailor their data sharing levels, addressing the one-size-fits-all privacy issue in federated recommender systems. Thirdly, to enhance model convergence speed, we propose an automatic client selection mechanism for federated recommendation systems, enabling the selection of optimal clients for training in each round to accelerate convergence. Finally, addressing the deployment challenges of federated recommender systems, particularly the significant parameter volume of embedding tables, we present a single-shot embedding dimension search method. This technique maintains model performance while reducing the size of embedding tables, facilitating efficient deployment on devices. Through these contributions, this thesis advances the field of federated graph neural network-based recommender systems, offering solutions to enhance accuracy, user control, and convergence speed, while ensuring privacy and reducing storage demands on local devices.

Bio: Mr. Liang Qu received a B.E. in Applied Physics in 2017, and an M.S. in Computer Science in 2019, from South China University of Technology and Harbin Institute of Technology respectively. Then he started his Ph.D. in computer science, under the supervision of Prof. Hongzhi Yin and Dr. Miao Xu. His main research interests are federated learning and recommender systems.