ChronoSAGE: Diversifying Topic Modeling Chronologically

In this paper, we propose a new chronological modeling of topics latent in documents. We apply sparse additive generative models (SAGE) [5] in a manner so that we diversify topic modeling results chronologically by using document timestamps. We call our approach ChronoSAGE. SAGE can represent each word probability by exponential of the sum of multiple parameters representing various facets of documents. Therefore, we prepare three types of parameter to utilize document timestamps: the parameters for each topic, those for each timestamp, and those for each pair of topic and timestamp. Consequently, word tokens are generated not only in a topic-specific manner, but also in a time-specific manner. We first compare ChronoSAGE and vanilla SAGE with LDA in terms of pointwise mutual information (PMI) [10] to show the practical effectiveness of SAGE-type approaches. We then give examples of time-differentiated latent topics obtained by ChronoSAGE to show the usefulness of our chronological topic modeling. As another contribution, we also provide an approximated inference that makes the implementation far easier.