Link prediction for heterogeneous information networks based on enhanced meta-path aggregation and attention mechanism

Heterogeneous link prediction aims to reveal potential connections between two nodes in heterogeneous information networks. Most existing studies are based on meta-paths, but ignore the information contained in incomplete meta-paths. They simply aggregate meta-paths, leading to mining semantic information insufficiently. To solve this problem, we propose a link prediction model based on enhanced meta-path aggregation and attention mechanism. In this model, the deficiency of missing topological information from incomplete meta-paths is compensated by aggregating structural features and semantics. Different from existing meta-path encoders, we use recurrent neural networks and the attention mechanism to learn explicit and implicit semantic knowledge from meta-paths, which can capture more complex semantic associations between nodes. In addition, to avoid duplicate feature acquisition by random walking, we design a novel bidirectional biased random walking algorithm. It is applied to guide the generation of heterogeneous neighbors of each node that contain features ignored by the meta-path-wise model, which can mine complete topological information and get more accurate link prediction results. The extensive experiments on several datasets demonstrate that the proposed model outperforms baselines.

Automated summary

Heterogeneous link prediction aims to reveal potential connections between nodes in heterogeneous networks. Existing studies based on meta-paths ignore the information in incomplete meta-paths, leading to insufficient mining of semantic information. To solve this problem, we propose a model that compensates for the deficiency of incomplete meta-paths by aggregating structural features and semantics. We use recurrent neural networks and attention mechanism to learn explicit and implicit semantic knowledge and design a bidirectional biased random walking algorithm to acquire complete topological information. The proposed model outperforms baselines in extensive experiments on multiple datasets.