Heterogeneous information network embedding based on multiperspective metapath for question routing

Abstract

Question routing is an important task on community question answering websites. Network embedding methods have achieved great success in question routing. However, most network embedding methods focus on the content of the questions and the answering history of the answerers. Answers that are similar to the accepted answer are treated in the same way as bad answers only because they are not “accepted”. As a result, the profiles of users are not fully assessed in these methods. To solve these problems, we propose a multiperspective metapath-based representation learning network for question routing, namely, MPQR. (1) MPQR learns multiperspective representations of question answerers, question raisers and questions based on a heterogeneous information network (HIN) that is constructed from answering records and voting information. Interest and expertise representations of users are learned at the same time. (2) A scoring function outputs the probability of each answerer providing the best answer. Pointwise loss and pairwise loss are combined to rank the answerers. The pointwise loss helps MPQR give more attention to potential answers with higher numbers of votes than bad answers. Experiments on real-world datasets show that MPQR outperforms state-of-the-art network embedding methods that ignore voting information.