Contrastive Topic Evolution Discovery via Nonnegative Matrix Factorization

In social networks, most information is stored, displayed and propagated in the form of text which is characterized by latent topics. With the discovery of latent topics, we can make more accurate prediction, inference and therefore, better applications, in social networks. Considering the dynamics and noise in social networks, we analyze data evolution from different perspectives including both common and discriminative topics. We adopt a basic model to factorize data matrix into components with an order representing latent topic and apply non-negative matrix factorization to two datasets as contrast to each other simultaneously. Time is taken into consideration as one order of the data matrix. Our basic model converts the decomposition problem into an optimization problem solved by block coordinate gradient descent, based on which the penalized model separates common part and discriminative part by adding penalty items. Next, a Bayesian generative model is established to avoid parameter tuning and overfitting, with the model parameters estimated via Gibbs sampling. Finally, we apply the models to both synthesized data and contrastive corpora of papers accepted by different conferences, as well as provide analysis of topic variation along with time.