Time Label Topic Model

Abstract

Most of the models not aware of these dependencies on document time stamps. Not modeling time can confound co-occurrence patters and results in exchangeability of topic problem, which is important factor to deal with when finding dynamic topic discovery. This limitation has thus motivated work on developing a generalized framework for incorporating time information into topic models. Consequently, a topic model named Topics over Time (TOT) is proposed, which introduces a time node in topic model to handle the exchangeability of topics problem. However it lacks the capability to accommodate data type of side information. In this paper, we present a generative time LDA-style topic model with a variety of side information named Time Label Topic(TLT), which can find not only how the latent low-dimensional structure of document-response pairs changes over time, but also overcome the exchangeability of topics problem. Empirical results demonstrate significant improvements accuracy of time stamp and response variable prediction, and lower perplexity of our proposed model and dominance over other models.