A Heterogeneous Network Representation Learning Approach for Academic Behavior Prediction

Abstract

Predicting authors’ academic behavior (e.g. co-authorship, citation) based on heterogeneous academic network can help scholars to grasp interesting research directions and participate in various co-operations. Most of the existing network representation methods use the structural and content features of nodes, but have not fully exploited the edges (relationships) between nodes (entities) and investigated the semantic compatibility of different edge types yet. To solve the above problems, a heterogeneous network representation learning method (HNEABP) is proposed to improve feature extraction and academic behavior prediction performance. HNEABP has three strengths: 1) capture rich neighbor information via balanced sampling and Skip-Gram, 2) apply knowledge graph embedding (KGE) technique to learn pairwise node information and to weight the importance of first-order neighbors, 3) solve the semantic incompatibility of edges based on KGE. Validation experiments on three academic network datasets show that HNEABP outperforms the popular network representation methods, which gives the credit to HNEABP for learning richer feature information effectively, so as to improve the performance of academic behavior prediction.