Modeling the evolution of associated data

• DOI: 10.1016/j.datak.2010.03.009
• Publication Date: 2010-03-26T10:29:24Z
• Authors: Jie Tang; Jing Zhang
• Keywords: Perplexity; Gibbs sampling; Baseline (sea)

Statistical topic models have been proposed for modeling documents and authorship information. However, few previous works have studied the evolution of associated data. In this paper, we investigate how to model trends of changes in document content and author interests simultaneously over time. We propose two models: a bag-of-words based Author-Time-Topic model that extends the state-of-the-art LDA-style topic model and a Hidden Markov Author-Time-Topic model, which can model interdependencies between topics. We use the Gibbs EM algorithm for parameter estimation. We apply these models to two data sets: NIPS papers and Yahoo group posts. Experimental results show that our models can achieve a lower perplexity (-2.0%-20%) than the baseline LDA and Author-Topic model, when modeling quickly evolving associated data. Experiments also reveal that the proposed models can accurately capture the hot topics in different periods (e.g. ''Yao at preseason'' in Aug-2004, when the Chinese player Ming Yao became a highlight in the NBA) from the two data sets.