GCN-based Explainable Recommendation using a Knowledge Graph and a Language Model

In this paper, we propose a novel graph convolutional network (GCN)-based recommendation method using both knowledge graph (KG) and review texts to solve the cold start problem and provide explainability. In our model, GCN-based collaborative filtering (CF) is parallelly performed on the user-item interaction graph and the KG to utilize the items’ additional information in the recommendation process. Also, we use a pretrained language model to generate embeddings from the user reviews and utilize them in the GCN embedding propagation process to reflect the users’ subjective sentiment and opinion. After the recommendation is performed, we generate the paths from the target user to recommended items by using the KG and embeddings from review texts. To prove the superiority of the proposed method, we conduct the experiment by comparing recommendation performance with baseline models. In the experiment, the proposed method outperformed the other models.