|
Abstract
|
Recommender systems have played a vital role in digital platforms, social networks, and content delivery services in recent years. Despite significant progress in Graph Neural Network (GNN)–based models and Contrastive Learning (CL), several fundamental challenges still hinder the delivery of accurate, stable, and fair recommendations. Major obstacles include data sparsity, the inherent noise in social and interactional relations, popularity bias, and the inability of existing models to effectively leverage community structures embedded within user networks. A key limitation of current methods is that they typically rely exclusively on either data-based augmentation or model-based view creation, while each alone is insufficient for producing the diversity and richness required for effective contrastive learning. Data-based methods often generate simple yet inadequate views through random edge deletion or addition, whereas model-based methods, even though capable of producing powerful views, may become unstable or overfit when they fail to align with the true structure of the underlying graph. This gap has resulted in the underutilization of critical information, particularly user community–related knowledge, across many existing models. The significance of community structures has grown substantially with the increasing integration of social recommender systems, as users in social networks naturally form communities characterized by shared interests and behavioral similarity. Leveraging community information not only enhances the quality of user and item representations but is especially crucial under cold-start and sparse conditions. However, most existing contrastive approaches lack mechanisms for generating community-aware views, thereby losing access to important latent information embedded within the social graph. To address this gap, the present research introduces a hybrid and modern framework for view creation in contrastive learning that synergistically com
|