Take a Further Step Beyond Data Augmentation: Augmented Knowledge Purification for Federated Recommendation

Abstract

By providing recommendation services while preserving user data privacy, federated sequential recommendation (FSR) achieves great attention recently. However, FSR suffers from the severe challenge of data sparsity, where item interaction sequences of each user are typically short. The common practice to address the problem is to employ various data augmentation operators such as crop or replace to generate more sequences. While generally adopted, we identify that these operators will bring additional noise, leading to unstable performance. To this end, this paper proposes FR-AKP, which performs the augmented knowledge purification on the basis of data augmentation at both the intra-client and inter-client levels. At the intra-client level, each client first applies augmentation operators to local sparse data, then transforms them into the frequency domain via Fast Fourier Transform, and finally employs a personalized frequency selection mechanism to denoise the augmented data. At the inter-client level, to prevent the propagation of noisy knowledge among clients during global aggregation, the server further adopts a fusion-network-based aggregation mechanism that generates personalized aggregation weights based on the distributional differences of the frequency-domain features of clients. Experimental results on four real-world datasets demonstrate that the proposed method significantly outperforms state-of-the-art methods by up to 34.62% in Recall and 29.92% in NDCG.