Modeling Topical Trends over Continuous Time with Priors

In this paper, we propose a new method for topical trend analysis. We model topical trends by per-topic Beta distributions as in Topics over Time (TOT), proposed as an extension of latent Dirichlet allocation (LDA). However, TOT is likely to overfit to timestamp data in extracting latent topics. Therefore, we apply prior distributions to Beta distributions in TOT. Since Beta distribution has no conjugate prior, we devise a trick, where we set one among the two parameters of each per-topic Beta distribution to one based on a Bernoulli trial and apply Gamma distribution as a conjugate prior. Consequently, we can marginalize out the parameters of Beta distributions and thus treat timestamp data in a Bayesian fashion. In the evaluation experiment, we compare our method with LDA and TOT in link detection task on TDT4 dataset. We use word predictive probabilities as term weights and estimate document similarities by using those weights in a TFIDF-like scheme. The results show that our method achieves a moderate fitting to timestamp data.