Personalizing Forum Search using Multidimensional Random Walks

Abstract

Online forums are a vital resource for users to ask questions and to participate in discussions. Yet, the search functionality on such forum sites is very primitive; posts containing the searched keywords are retrieved in the order of their creation date. In these interactive and social web forum sites, users frequently make connections with other users due to shared interests, same information needs or similar profiles. A critical challenge then, is to score and rank the forum posts while taking into account these relations between users. In this paper, we present a personalized search over forums that leverages user similarities developed via multiple relations linking users. We build a novel multidimensional random walk model that uniformly incorporates the heterogeneous user relations to find similar forum participants. We then use this multi-relational user similarity to predict future interactions by personalizing answer search. Furthermore, we extend our methods to enhance keyword search for forum readers, by using expertise scores for all existing forum participants. Our results show that by leveraging the author dimension we can retrieve more relevant results than the traditional IR scoring alone.