Learning for Expertise Matching with Declination Prediction

We study the problem of finding appropriate experts who are able to complete timely reviews and would not say “no” to the invitation. The problem is a central issue in many question-and-answer systems, but has received little research attention. Different from most existing studies that focus on expertise matching, we want to further predict the expert’s response: given a question, how can we find the expert who is able to provide a quality review and will agree to do it. We formalize the problem as a ranking problem. We first present an embedding-based question-to-expert distance metric for expertise matching and propose a ranking factor graph (RankFG) model to predict expert response. For online evaluation, we developed a Chrome Extension for reviewer recommendation and deployed it in the Google Chrome Web Store, and then collected the reviewers’ feedback. We also used the review bidding of a CS conference for evaluation. In the experiments, the proposed method demonstrates its superiority (+6.6-21.2% by MAP) over several state-of-the-art algorithms.