SocialKAN: Learning User Preferences on Signed Social Graphs with Kolmogorov-Arnold Networks
Abstract
User-user social networks are widely utilized as side information in recommendation and have been shown to improve performance. We argue that current social recommenders still have room for improvement. First, for social data utilization, most methods focus mainly on positive trust relations, while failing to effectively use negative distrust relations, which can also provide useful signals for recommendation. Second, in terms of modeling architecture, most methods rely on linear or nonlinear transformation structures (e.g., multi-layer perceptrons) as their foundational building blocks. Kolmogorov-Arnold Networks (KANs), recently introduced for their strong data-fitting ability, offer a promising alternative. To this end, in this paper, we propose SocialKAN, a novel KAN-based trust-distrust-aware social recommender method. Specifically, SocialKAN introduces three KAN-based graph learning channels to model the user-item interaction graph, user trust graph, and user distrust graph, respectively. To better enable effective cross-view adaptation among the three channels, we design a semantic routing module for knowledge sharing at the embedding level and an adaptive consensus knowledge distillation strategy for output alignment at the optimization level. Experimental results on the three datasets show that our proposed SocialKAN method achieves strong performance. Ablation studies further validate the rationality and effectiveness of the components designed in our SocialKAN.