Social Network Based Recommender System: A Probabilistic Graphical Model Approach

The internet presents a vast labyrinth of information and choices, overwhelming users while service providers strive to predict user preferences accurately. Traditional recommender systems, although utilizing extensive data, often lack the social context necessary for precise recommendations, leading to issues like cold-start and data sparsity. Social network-based recommender systems have been proposed to address these limitations by leveraging the decision-making processes influenced by social connections. Recent advancements in this domain, such as the Diffnet and Diffnet++ models, have shown improved prediction accuracy through deep social influence propagation models. However, their performance diminishes with higher diffusion depths. In this work, we propose enhancements to the Diffnet and Diffnet++ models by incorporating the homophily effect in the heterogeneous graph of the diffusion layer and implementing a weighted embedding method in the prediction layer. Using data from Yelp and Flickr, our improved model integrates higher-order social structures and interest network information, demonstrating superior performance on metrics such as Normalized Discounted Cumulative Gain (NDCG) and Hit Ratio (HR) compared to existing baselines. Our approach effectively integrates social network data and user similarity networks, resulting in more accurate and reliable recommendations for online platforms where social networking is a key feature.