Knowledge Graph Double Interaction Graph Neural Network for Recommendation Algorithm

Abstract

To solve the problem that recommendation algorithms based on knowledge graph ignore the information of the entity itself and the user information during information aggregating, we propose a double interaction graph neural network recommendation algorithm based on knowledge graph. First, items in the dataset are selected as user-related items and then they are integrated into user features, which are enriched. Then, according to different user relationship weights and the influence weights of neighbor entities on the central entity, the graph neural network is used to integrate the features of nodes in the knowledge graph to obtain neighborhood information. Secondly, user features are interacted and aggregated with the entity’s own information and neighborhood information, respectively. Finally, the label propagation algorithm is used to train the edge weights to assist entity features learning. Experiments on two real datasets commonly used in recommended algorithms were conducted and showed that the model is better than the existing baseline models. The values of AUC and F1 on MoviesLens-1M are 0.905 and 0.835 and on the Book-Crossing are 0.698 and 0.640. Compared with the baseline model, the Precision@K index improved by 1.3–3% and the Recall@K index improved by 2.2~11.2% on the MoviesLens-1M dataset, while the Precision@K index improved by 0.6~1.6% and the Recall@K index improved by 4.5~10.8% on the Book-Crossing dataset. The model also achieves strong performance in data-sparse scenarios.

Keywords

• recommender systems
• knowledge graph
• graph neural networks
• personalized recommendation