Abstract
Social recommender systems help address data sparsity in user–product interactions by leveraging social relationships to infer user preferences. However, existing models often overlook the role of social capital that influence decision-making in social commerce. Social capital consists of structural, relational, and cognitive dimensions, all of which shape user preferences. To better understand these influences, we propose a multi-task learning framework named DeepSC that integrates social capital theory into preference modeling. Its user preference learning module extracts structural features through graph-based pre-training, learns relational features from dynamic user embeddings, and models cognitive features using a hypergraph attention network. Additionally, the dual graph-based product feature learning module enhances cognitive feature extraction by incorporating product co-interactions. DeepSC is optimized through a joint learning objective, combining point-wise and pair-wise learning with an auxiliary social link prediction task to refine user representations. Experiments on three e-commerce datasets demonstrate that DeepSC significantly outperforms the state-of-the-art recommendation models, highlighting the effectiveness of integrating social capital into social preference learning. Our research advances social recommendation by providing a social capital theory-driven approach to modeling user behavior in digital commerce.