Improved topic models for social media via community detection using user interaction and content similarity

Topic models such as Latent Dirichlet Allocation(LDA) have historically served as a successful tool for various data mining applications on conventional documents such as news articles or academic abstracts. However, standard use of topic models on social media posts pose several poblems because social media posts are short, messy and generated non-uniformly by the users of the social media platforms. In this paper we propose a new approach of community based document pooling to train better topic models over social media posts and address these problems without modifying the basic machinery of LDA. We compare our approach to the popular user based pooling scheme and show significant improvement in the quality of topic models.