Link Prediction Based on Deep Global Information in Heterogeneous Graph

Link prediction on heterogeneous graph is helpful to improve real network information, promote the development of recommendation system and explore potential correlation. Some existing models and methods have proved that mining the structure and semantic information of heterogeneous graph can bring better results, but these methods can not extract the information of heterogeneous graph in a deeper level, or can not widely and effectively disseminate node information, resulting a poor performance on some complex heterogeneous graph. Therefore, a link prediction model KMHLP based on global structure information and multi-level semantic information is proposed in this paper. On the one hand, this paper proposes RBC, a node importance index that can deeply mine graph structure information, and uses RBC matrix to improve GCN to propose K-GCN layer, and then uses K-GCN layer to build KHLP module, experiments show that this module can better extract information representation containing global structure. On the other hand, in order to better refine node characteristics, we use attention mechanism to collect the semantic information contributions of different metapath instances to nodes under the same type of metapath. In addition, we consider the impact of different types of metapath on the role information of the central node, and construct cross-metapath type attention mechanism. Finally, the KMHLP model realizes the enhancement of the effect again through the effective feature aggregation. In this paper, a large number of comparison and ablation experiments were conducted on two datasets, IMDB and LastFM. The experimental results show that KMHLP achieves the best performance in the link prediction task of IMDB and LastFM. Our code has been open-sourced: https://github. com/776166320/KMHLP.