Feature Similarity Learning Enhanced Knowledge Graph-based Convolutional Networks for Recommendation

• Published in: 2021 6th International Symposium on Computer and Information Processing Technology (ISCIPT) pp. 693 - 698
• Main Authors: Shi, Boya, Zheng, Zhong, Tian, Tao, Du, Wanru
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2021
• Subjects: Benchmark testing; Computer architecture; Graph convolutional networkscomponent; Graph neural networks; Knowledge engineering; Knowledge Graph; Multi-Head Attention; Recommender Systems
• DOI: 10.1109/ISCIPT53667.2021.00146

In the recommendation algorithm based on the knowledge graph, the introduction of graph neural network effectively mines the rich auxiliary information in knowledge graph. While exploring the structure of graph neural network, more studies have begun to strengthen the mining of user historical interactive information. However, these methods ignore the similarity of user characteristics, so the learned embedding representation fails to fully reflect user preferences. In this article, we propose a new model FSKGCN. By modeling the similarities between user characteristics, we design an architecture that integrates and knowledge propagation and user similar characteristics. Specifically, FSKGCN develops a user feature similarity modeling layer, which calculates the similarity between user features by introducing a multi-head attention mechanism, and further encodes the user's high-order features according to the similarity score. Compared with other KG-based methods, FSKGCN provides a new architecture that combines knowledge information with user similar characteristics. We conducted CTR experiments on two real-world datasets, and the results show that FSKGCN is significantly better than the latest benchmark method.