Collaborative Recommendation of Temporally-Discounted Tag-Based Expertise for Community Question Answering

Abstract

We propose an innovative approach to finding experts for community question answering (CQA). The idea is to recommend answerers, who are credited the highest expertise under question tags at routing time. The expertise of answerers under already replied question tags is intuitively discounted by accounting for the observed tags, votes and temporal information of their answers. Instead, the discounted expertise under not yet replied tags is predicted via a latent-factor representation of both answerers and tags. These representations are inferred by means of Gibbs sampling under a new Bayesian probabilistic model of discounted user expertise and asking-answering behavior. The devised model unprecedentedly explains the latter two CQA aspects as the result of a generative process, that seamlessly integrates probabilistic matrix factorization and network behavior characterization. An extensive comparative experimentation over real-world CQA data demonstrates that our approach outperforms several-state-of-the-art competitors in recommendation effectiveness.