Multi-semantic and multi-behavior recommendation with graph attention networks

Recommender systems have become an essential tool to help resolve the information overload problem in recent years. Traditional recommender systems usually rely on Collaborative Filtering (CF) methods based on user-item interactions. However, these methods face data sparsity issues in most cases, which negatively impact recommendation performance and reduce user experience. Recent researches focus on content enriched methods incorporating semantic relevance enhancing the representation learning and improving data utilization. To this end, this paper proposes a recommendation model, named MSMB-GAN (Multi-Semantic Multi-Behavioral Graph Attention Network) aimed at improving recommendation accuracy. To construct this model, multi-facet information is made full use, including user-item interaction behavior (like view, click, add to cart, etc.), social network, and item category information. An encoder is used to capture different behavior relationship between users and items; and a graph attention mechanism is used to capture the strength of user-user social relationship; and a decay factor is introduced to represent the changing demand of users to different item categories in this model. Finally, the user representations extracted from the user-item interaction behaviors, user-user social relationships, and item-category correspondence information are weighted and integrated. Experimental results on synthetic and real-world datasets show that MSMB-GAN can achieve better recommendation performance compared to some existing methods. Further ablation experiments verify the effectiveness of utilizing multi-source information.