Detecting Social Topic by Hashtag-Weighted Topic Model over Time

Nowadays, more and more social media platforms support hashtags to facilitate information classification. Like Twitter hashtags, a user-initiated hashtag can suggest emotion/mood, convey so much extra information in addition to the actual tweet. Hashtags have been widely used in topic analysis because of its informative effect, but all hashtags are created equally. In the paper, we propose a Hashtag-Weighted Topic Model over Time (HWOT) which assigns hashtags to deal with topic evolving over time with different hashtag weight. To leverage hashtags across topics in a specific time period, the topic of hashtag is represented as a multinomial distribution and the topic over time as a Beta distribution. Our model can uncover the latent relationships among topics, hashtags and time. The weight of the hashtag is learned via a novel context aware weakly supervised approach. Experiments on Twitter dataset show that our model can achieve better performance in terms of model perplexity. It further reveals the change of the topics over time.