Graph learning representation for item recommendation

Abstract

In recent years, recommender systems have been well developed and have been applied to many aspects such as product recommendation in e-commerce platforms and friend recommendation in social media. Moreover, graph representation learning methods are adopted into recommender systems and achieve good performance. Traditional methods pay more attention to the interaction between users and items while ignoring the side information of items, which may lead to information loss. Knowledge Graph Attention Network (KGAT) deploys recommendation problems to knowledge graphs which have the ability to record both user-item interactions and high-order relations. Node embeddings are generated by recursively propagating its neighbors using an attention mechanism. However, the original methods take every neighbor into account, which is very time-consuming. Therefore, we propose a new embedding layer with a random walk based aggregator to sample the neighbors and aggregate them by importance score. Experiments on three benchmark datasets are conducted, and our proposed method achieves comparable performance accuracy with a faster time.