Expert recommendation is a central task in question-answering communities that involves facing several issues, such as quantifying users' expertise and replying propensity, discounting both over time, and taking advantage of the social graph of cooperative asking-answering interactions. In this paper, we propose two model-based approaches to recommend repliers in question-answering communities, whose novelty lies in the above issues being tackled jointly. The devised approaches focus on tags to avoid processing large amounts of (long and short) text messages.The first approach routes questions to answerers with the greatest expertise in the topics marked by question tags at routing time. To this end, the expertise of community members under such tags is distinguished between shown and unknown. More precisely, the expertise shown by answerers in previously-covered question topics is discounted by looking at the temporal information, votes, and tags of their past answers. Instead, the discounted degree of the unknown expertise of answerers in not-yet-covered question topics is predicted through the embeddings of tags and answerers into a space of latent factors. Such embeddings stem from Gibbs sampling inference under a fully-Bayesian latent-factor model of the temporally-discounted tag-based expertise of users and the social graph of their replying behavior.The second approach recommends repliers by also accounting for their temporally-discounted, tag-based answering propensity. In essence, a higher propensity to answer questions with certain tags is more likely to be expected in those community members who have answered a larger number of earlier questions with those tags, especially in the recent past.An extensive experimentation with real-world benchmark datasets demonstrates the superiority of our approaches over state-of-the-art competitors in recommendation effectiveness, according to several evaluation criteria. In particular, the average competitive gain of the first and second approaches is 11% and 13%, respectively, across the datasets.
Here are the answers. What is your question? Bayesian collaborative tag-based recommendation of time-sensitive expertise in question-answering communities
Riccardo Ortale
2023
Abstract
Expert recommendation is a central task in question-answering communities that involves facing several issues, such as quantifying users' expertise and replying propensity, discounting both over time, and taking advantage of the social graph of cooperative asking-answering interactions. In this paper, we propose two model-based approaches to recommend repliers in question-answering communities, whose novelty lies in the above issues being tackled jointly. The devised approaches focus on tags to avoid processing large amounts of (long and short) text messages.The first approach routes questions to answerers with the greatest expertise in the topics marked by question tags at routing time. To this end, the expertise of community members under such tags is distinguished between shown and unknown. More precisely, the expertise shown by answerers in previously-covered question topics is discounted by looking at the temporal information, votes, and tags of their past answers. Instead, the discounted degree of the unknown expertise of answerers in not-yet-covered question topics is predicted through the embeddings of tags and answerers into a space of latent factors. Such embeddings stem from Gibbs sampling inference under a fully-Bayesian latent-factor model of the temporally-discounted tag-based expertise of users and the social graph of their replying behavior.The second approach recommends repliers by also accounting for their temporally-discounted, tag-based answering propensity. In essence, a higher propensity to answer questions with certain tags is more likely to be expected in those community members who have answered a larger number of earlier questions with those tags, especially in the recent past.An extensive experimentation with real-world benchmark datasets demonstrates the superiority of our approaches over state-of-the-art competitors in recommendation effectiveness, according to several evaluation criteria. In particular, the average competitive gain of the first and second approaches is 11% and 13%, respectively, across the datasets.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.