Interaction Graph Neural Network for News Recommendation

摘要

Personalized news recommendation has become a highly challenging problem in recent years. Traditional ID-based methods such as collaborative filtering are not suitable for news recommendation due to the extremely rapid update of candidate news. Various content-based methods have been proposed for news recommendation and achieved the state-of-the-art performance. Recently, knowledge-aware news recommendation further improves the performance through discover latent knowledge level connections among the news. However, we argue that the above content-based methods do not fully utilize the collaborative information latent in user-item interactions into user and news representation learning process. In this paper, we propose a new news recommendation model, Interaction Graph Neural Network (IGNN), which integrates a user-item interactions graph and a knowledge graph into the news recommendation model. Specifically, IGNN obtains the representation of users and items with two graphs. One is the knowledge graph, and another is the user-item interaction graph. It learns the content-based feature from knowledge-level and semantic-level with convolutional neural networks and fuses the high-order collaborative signals extracted from the user-item interaction graph into user and news representation learning process with a graph neural network. Extensive experiments are conducted on the two real-world news data sets, and experimental results show that IGNN significantly outperforms the state-of-the-art approaches for news recommendation.