Steering Time-Dependent Estimation of Posteriors with Hyperparameter Indexing in Bayesian Topic Models

Abstract

This paper provides a new approach to topical trend analysis. Our aim is to improve the generalization power of latent Dirichlet allocation (LDA) by using document timestamps. Many previous works model topical trends by making latent topic distributions time-dependent. We propose a straightforward approach by preparing a different word multinomial distribution for each time point. Since this approach increases the number of parameters, overfitting becomes a critical issue. Our contribution to this issue is two-fold. First, we propose an effective way of defining Dirichlet priors over the word multinomials. Second, we propose a special scheduling of variational Bayesian (VB) inference. Comprehensive experiments with six datasets prove that our approach can improve LDA and also Topics over Time, a well-known variant of LDA, in terms of test data perplexity in the framework of VB inference.