WCSE

Transformer Recommendation Algorithm with Fusion Knowledge Graph

Abstract— To address the problem that traditional knowledge graph recommendation algorithms only consider the message aggregation of neighbor nodes but not the sequence information of knowledge graph neighbor nodes, this paper proposes a Transformer recommendation algorithm that fuses knowledge graphs. Under the item knowledge graph space, the sequence information is constructed from the target item and its neighbors, and the sequence information is encoded using the Transformer model to capture the features of the item. Finally, MLP is used to integrate the obtained item feature vectors and user embedding vectors and send them to the prediction module to obtain user click-through rate prediction by vector inner product operation. Comparast experiments were performed on three publicly available datasets. On the MovieLens- 20M dataset, compared with the optimal baseline KGCN, the AUC and F1 improve by 1.74% and 7.18%, respectively. On the Last.FM dataset, the AUC and F1 improve by 7.07% and 5.92%, respectively. On the Book-Crossing dataset, the AUC and F1 improve by 1.52% and 1.62%, respectively, compared to the optimal baseline. Thus, it is verified that the effectiveness of the algorithm is improved by considering the sequence information of the neighboring nodes of the knowledge graph.

Index Terms— Knowledge Graph, Transformer, Recommendation Algorithm.

Honghui Xie
School of Software, Jiangxi Agricultural University, China
Jun Yang
School of Software, Jiangxi Agricultural University, China
Yi Liu
School of Software, Jiangxi Agricultural University, China
Cong Huang
School of Software, Jiangxi Agricultural University, China

[Download]

Cite: Honghui Xie, Jun Yang, Yi Liu, Zhen Wang, " Transformer Recommendation Algorithm with Fusion Knowledge Graph," WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications, pp. 285-292, Sanya, China, April 15-18, 2022.