RIBGCL: Reliability-Guided Item-Bias-Aware Graph Contrastive Learning for Social Recommendation

Abstract Noisy social ties and popularity-skewed interactions can weaken graph-based social recommendation by transferring irrelevant preferences and overemphasizing frequently observed items. We propose RIBGCL (Reliability-Guided Item-Bias-Aware Graph Contrastive Learning), a framework that coordinates confidence-weighted social propagation with a bounded item-representation correction. First, an interest-consistency gate estimates the relative compatibility of socially connected users and reweights their messages while retaining residual social connectivity. Second, an item-neighborhood branch derives a normalized correction direction and applies similarity-controlled, weak residual refinement to item embeddings. These two components target complementary sources of representation contamination and are trained jointly with social and interaction objectives and augmentation-free contrastive learning. Analytically, the gate preserves a positive lower bound on observed social-edge weights, while the unit-normalized item correction bounds the perturbation of each predicted score. On Epinions, Ciao, and FilmTrust, RIBGCL improves the reported Top-10 ranking metrics over the evaluated baselines; component ablations and social-edge filtering comparisons support the usefulness of both modules. Because the item correction is derived from observational interactions, it is treated as counterfactual-inspired rather than as an identified causal effect.