Efficient yet secure: An archive knowledge graph-enhanced native sparse attention network for lightweight privacy-preserving recommendation

Abstract

Recommendation Systems (RSs) aim to provide personalized recommendations by modeling user-item interaction patterns. Current attribute-enhanced RSs leverage user archival attributes to improve predictive performance. However, the use of attribute information introduces two critical challenges: 1) the risk of privacy leakage, as sensitive user attributes can be inferred from learned representations, and 2) high computational complexity, primarily due to the quadratic complexity of attention mechanisms. To address the accuracy-privacy-efficiency trilemma, we propose an Archive Knowledge Graph-enhanced Native Sparse Attention network (AKG-NSA) for privacy-preserving lightweight recommendation. Specifically, AKG-NSA introduces a two-stage privacy protection mechanism. First, we pseudonymize user identities in the archive knowledge graph, breaking the direct linkage between users and their attributes. Second, we design a Multi-channel Native Sparse Attention (MNSA) network that utilizes compressed user representations as queries to retrieve attribute patterns from the archive knowledge graph in a privacy-preserved manner. Moreover, we also construct a parallel user-item bipartite graph and operate graph convolutions to learn the representations for users and items. By employing the native sparse attention mechanism, AKG-NSA refines the learned representations while maintaining a low computational complexity. Extensive experiments on three real-world datasets demonstrate that AKG-NSA outperforms nine state-of-the-art baselines in terms of prediction accuracy, privacy preservation, and computational efficiency. The data and source codes of this work are available at https://github.com/juandu113/AKG-NSA.