Knowledge graph enhanced multi-task learning between reviews and ratings for movie recommendation
Yun Dong Liu, Jun Miyazaki, Qiong Chang · Proceedings of the 37th ACM/SIGAPP Symposium on Applied Computing · 2022
In many current recommender systems, online reviews are used to boost the recommendation performance. As historical ratings and reviews are the two main instances of user feedback, and the combination of these two is very important to understand why a user likes or dislikes an item. A general way of learning the correlation between reviews and ratings is to learn their distributions in the same latent semantic space, where each dimension is a set of item features. However, because there are only a few words related to items in reviews, this method may introduce noise and output misunderstandings. In addition, it is difficult to align heterogeneous data using general ways. In this paper, we introduce a knowledge graph enhanced multi-task learning method to learn cross features from ratings and reviews by fusing users and movies with their review knowledge entities in the same graph. The fusion graph structure enables the graph-link prediction task to learn the review entity features that are relevant to target users and movies, which further assists our recommendation task. Experiments on 18 different sparse datasets that were preprocessed from two public datasets demonstrate the effectiveness of our model. Our method also achieves good performance when user-movie interactions are sparse.