Rethinking the Role of Social Signal in Graph-Based Recommendation

Abstract

Social recommendation models integrate user friendship networks into collaborative filtering under the assumption that connected users share preferences. These models are typically evaluated on static benchmarks where network topology, interaction density, and homophily are confounded, making it difficult to identify when social signal integration actually helps. This thesis addresses the gap with a controlled synthetic benchmark whose graph topology, signal quality, and structural integrity can be varied independently, along with validation on three real-world datasets. We evaluate five established models covering standard collaborative filtering baselines, contrastive self-supervision, and social-aware paradigms (pairwise diffusion and higher-order motif aggregation). We also propose three lightweight hybrid architectures that integrate social signal into a LightGCN framework through different gating mechanisms. Our results show that architectural choices dominate graph topology across all conditions tested, and that the computational overhead of social integration does not produce proportional accuracy returns.