Link prediction in heterogeneous information networks: An improved deep graph convolution approach

Abstract

Heterogeneous information networks (HINs) refer to logical networks involving entities of multiple types and their multiple relations, which are widely used for modeling real-world systems with rich features and intricate patterns. Link prediction in such networks is a consistent interesting research question due to its methodological and practical implications in the business field. This study develops an improved spatial graph convolution network to learn predictive vertex embeddings with minimal information loss based on local community discovery and to handle the complexity of link predictions in the context of HINs. An optimizable kernel layer is designed to measure the similarity of pairwise vertex embeddings. The effectiveness of the proposed method is validated using four real-world HINs: WordNet, MovieLens, DBLP, and Douban. The results of the experiments demonstrate that the proposed method outperforms several benchmark algorithms, achieving F1-scores of 87.65%, 84.27%, 82.99%, and 89.96% on the four HINs, respectively. The findings of this study can inform the design and improvement of related information systems.