Heterogeneous Information Network enhanced Academic Paper Recommendation

Abstract

Academic paper recommender (APR) systems that assist researchers in solving the information overload problem have attracted lots of attention. Recently, many works have been done to improve APR with heterogeneous information network (HIN). However, these works plainly depend on graph embedding to generate recommendations and achieve unsatisfactory performance due to the neglect of high-order paper connectivity in the HIN and complex interactions between the user and academic papers. This paper proposes a new algorithm named Heterogeneous Information Network enhanced Academic Paper Recommendation (HIN-APR) to address the above problems. Firstly, based on the message-passing architecture of GNN, designing a novel heterogeneous graph neural network including dual-level attention is to learn the paper’s high-order feature in HIN. Then, the high-order feature was integrated into a new recommendation framework based on convolution neural network (CNN) to model the complex interactions and predict matching score between users and papers. Experimental results on citeulike-a and citeulike-t show that our proposed approach outperforms compared with baseline methods.