Time-Sensitive Topic-Based Communities on Twitter

This paper tackles the problem of detecting temporal content-based user communities from Twitter. Most existing content-based community detection methods consider the users who share similar topical interests to be like-minded and use this as a basis to group the users. However, such approaches overlook the potential temporality of users' interests. In this paper, we propose to identify time-sensitive topic-based communities of users who have similar temporal tendency with regards to their topics of interest. The identification of such communities provides the potential for improving the quality of community-level studies, such as personalized recommendations and marketing campaigns that are sensitive to time. To this end, we propose a graph-based framework that utilizes multivariate time series analysis to represent users' temporal behavior towards their topics of interest in order to identify like-minded users. Further, Topic over Time (TOT) topic model that jointly captures keyword co-occurrences and locality of those patterns over time is utilized to discover users' topics of interest. Experimental results on our Twitter dataset demonstrates the effectiveness of our proposed temporal approach in the context of personalized news recommendation and timestamp prediction compared to non-temporal community detection methods.