Detecting short-term cyclical topic dynamics in the user-generated content and news

With the maturation of the Internet and the mobile technology, Internet users are now able to produce and consume text data in different contexts. Linking the context to the text data can provide valuable information regarding users' activities and preferences, which are useful for decision support tasks such as market segmentation and product recommendation. To this end, previous studies have proposed to incorporate into topic models contextual information such as authors' identities and timestamps. Despite recent efforts to incorporate contextual information, few studies have focused on the short-term cyclical topic dynamics that connect the changes in topic occurrences to the time of day, the day of the week, and the day of the month. Short-term cyclical topic dynamics can both characterize the typical contexts to which a user is exposed at different occasions and identify user habits in specific contexts. Both abilities are essential for decision support tasks that are context dependent. To address this challenge, we present the Probit-Dirichlet hybrid allocation (PDHA) topic model, which incorporates a document's temporal features to capture a topic's short-term cyclical dynamics. A document's temporal features enter the topic model through the regression covariates of a multinomial-Probit-like structure that influences the prior topic distribution of individual tokens. By incorporating temporal features for monthly, weekly, and daily cyclical dynamics, PDHA is able to capture interesting short-term cyclical patterns that characterize topic dynamics. We developed an augmented Gibbs sampling algorithm for the non-Dirichlet-conjugate setting in PDHA. We then demonstrated the utility of PDHA using text collections from user generated content, newswires, and newspapers. Our experiments show that PDHA achieves higher hold-out likelihood values compared to baseline models, including latent Dirichlet allocation (LDA) and Dirichlet-multinomial regression (DMR). The temporal features for short-term cyclical dynamics and the novel model structure of PDHA both contribute to this performance advantage. The results suggest that PDHA is an attractive approach for decision support tasks involving text mining.