Graph Intention Embedding Neural Network for tag-aware recommendation.

Tag-aware recommender systems leverage the vast amount of available tag records to depict user profiles and item attributes precisely. Recently, many researchers have made efforts to improve the performance of tag-aware recommender systems by using deep neural networks. However, these approaches still have two key limitations that influence their ability to achieve more satisfactory results. Firstly, traditional approaches cannot fully exploit the rich content and associated features underlying the users' tagging history. Secondly, most existing models ignore the intention behind user-item interactions while extracting user interest from these behaviors, leading to a deficiency in interpretability. In this work, we have proposed a new model for tag-aware recommendation, namely Graph Intention Embedding Neural Network (GIENN), to address these two issues. Specifically, the proposed approach GIENN initially constructs a tag-aware interaction graph (TAIG) based on users' tagging history and then devises a two-layer attention mechanism to simultaneously learn the significance of both node neighbors and node types. Furthermore, GIENN leverages the complex semantic information inherent in tags to reveal the true intention behind user interactions. The information of tags is propagated to the embedding of other nodes (including users and items) in graph with attentive weights to indicate which tags are more influential in user-item interactions. Comprehensive experiments on three public datasets demonstrate that GIENN performs better than state-of-the-art baselines in tag-aware top-N recommendation tasks.