Ask and Ye shall be Answered: Bayesian tag-based collaborative recommendation of trustworthy experts over time in community question answering

Several challenging issues have yet to be jointly addressed in the recommendation of experts for community question answering, including dynamicity, comprehensive profiling, the incorporation of auxiliary data, and the manipulation of heterogeneous information. We argue that a unified treatment of these issues is beneficial in improving recommendation effectiveness.In this paper, we introduce and formalize a new and more thorough instance of the expert-recommendation task for community question answering, which is conceived to suitably account for the connections among the targeted issues. Moreover, in order to carry out the devised task, we present an innovative Bayesian tag-based approach that systematically handles all of the targeted issues in a coherent and seamlessly unified manner. At the heart of the proposed approach is an unprecedented principled fusion of various types of information. The integrated information enables a peculiar characterization of community members in terms of three inherent properties, i.e., their tag-based temporally-discounted interest, expertise, and willingness to respond. The first property is determined by looking into questions, while the last two are determined from answers. Within a generic question answering community, the three properties of its members allow for selectively routing any question to a specifically addressed set of responders. These are recommended as trustworthy repliers, who are not only actually experts in the themes of the routed question as per its tags, but also still interested in such themes and willing to answer at routing time.An extensive empirical assessment involving real-world benchmark data from heterogeneous domains reveals the higher recommendation effectiveness of the presented approach compared to state-of-the-art competitors.

Automated summary

There are several challenging issues in the expert recommendation for community question answering that needs to be addressed, such as dynamicity, comprehensive profiling, incorporation of auxiliary data, and manipulation of heterogeneous information. A unified treatment of these issues is believed to improve recommendation effectiveness.