Probabilistic topic model based approach for detecting bursty events from social media data

To detect bursty events from the huge amount of real-time data generated from various social networks has attracted more and more research efforts. Most of existing algorithms detect the bursty events either by discovering the co-occurrent bursty words or the emerging topics, ignoring the association between bursty and topics. Meanwhile, these algorithms are not able to cope with short text data like Weibo and Twitter. This paper proposes two novel probabilistic generative models (TBE/TBEP). TBE model can detect bursty events on long articles which can simultaneously consider the co-occurrent relationships among bursty words as well as the co-occurrent relationships among occurrent words and the underlying topics which generate the bursty events. TBEP model captures the assumption: one post are always have the one topic, which can handle the bursty events on Weibo and Twitter. The Gibbs sampling technique is adopted to estimate the model parameters. Extensive experiments are performed on three real data sets and the promising results, compared with the state-of-the-art Hot-Bursty-Event detection algorithms, have demonstrated that the proposed approach can: (1) achieve better model performance with respect to the evaluation criteria; (2) achieve more accurate bursty evnets on long/short text data.