Heterogeneous graph neural networks contrastive learning based recommender system

In recent years, the increase in activity on digital platforms has played a significant role in the growing interest in recommender systems. The diversification of user preferences challenges the limitations of traditional recommendation approaches, accelerating innovative research in this field. The success of Graph Neural Networks (GNN) in various application areas and the structural suitability of data in recommender systems have increased interest in GNN-based recommendation models. Despite successfully capturing high-order relations between users and items, these systems still face significant challenges such as data sparsity. As a solution to this problem, social recommendation methods that aim to improve recommendation performance by leveraging users' social networks are present in the literature. This thesis introduces an innovative heterogeneous structured model that integrates not only users' social relationships but also the categorical relationships of items. The model provides recommendations by utilizing the user-item graph, user-user social network graph, and item-category hypergraph with a simple and effective structure. Additionally, it enhances the effectiveness of the model by leveraging alignment between representations of different views through a contrastive learning mechanism. Comprehensive experiments demonstrate that the model outperforms baseline models. This study contributes significantly to the field of recommendation systems, highlighting the potential of heterogeneous structured graph neural networks and contrastive learning-based recommendation systems, offering new perspectives for future research.