HOFA: Higher-Order Feature Alignment for Social Recommendation via Contrastive Learning

In social recommendation, user relationships are often utilized to enhance personalized suggestions. However, performance is hindered by sparse social links and the cold-start problem in user-item interactions. While contrastive learning helps alleviate data sparsity by learning generalizable representations, existing methods still face key challenges: (1) limited exploitation of higher-order social features; (2) lack of mechanisms for feature uniformity, leading to redundant correlations; and (3) insufficient feature decorrelation, reducing representation discrimination. To address these issues, we propose HOFA (Higher-Order Feature Alignment for Social Recommendation via Contrastive Learning), a contrastive learning framework for social recommendation. HOFA captures informative higher-order structures via triangular motifs and aligns features on the hypersphere to promote uniformity and decorrelation. This reduces redundancy and enhances embedding discrimination. HOFA also optimizes the distribution of users and items in the embedding space and minimizes positive distances to better reflect user preferences. Extensive experiments on three publicly available real-world datasets demonstrate that HOFA significantly surpasses several representative methods in the Top-K recommendation task, underscoring its notable benefits.