Co-clustering for Federated Recommender System

Abstract

As data privacy and security attract increasing attention, Federated\nRecommender System (FRS) offers a solution that strikes a balance between\nproviding high-quality recommendations and preserving user privacy. However,\nthe presence of statistical heterogeneity in FRS, commonly observed due to\npersonalized decision-making patterns, can pose challenges. To address this\nissue and maximize the benefit of collaborative filtering (CF) in FRS, it is\nintuitive to consider clustering clients (users) as well as items into\ndifferent groups and learning group-specific models. Existing methods either\nresort to client clustering via user representations-risking privacy leakage,\nor employ classical clustering strategies on item embeddings or gradients,\nwhich we found are plagued by the curse of dimensionality. In this paper, we\ndelve into the inefficiencies of the K-Means method in client grouping,\nattributing failures due to the high dimensionality as well as data sparsity\noccurring in FRS, and propose CoFedRec, a novel Co-clustering Federated\nRecommendation mechanism, to address clients heterogeneity and enhance the\ncollaborative filtering within the federated framework. Specifically, the\nserver initially formulates an item membership from the client-provided item\nnetworks. Subsequently, clients are grouped regarding a specific item category\npicked from the item membership during each communication round, resulting in\nan intelligently aggregated group model. Meanwhile, to comprehensively capture\nthe global inter-relationships among items, we incorporate an additional\nsupervised contrastive learning term based on the server-side generated item\nmembership into the local training phase for each client. Extensive experiments\non four datasets are provided, which verify the effectiveness of the proposed\nCoFedRec.\n