A0320
Title: Time-varying Poisson factorization with an application to U.S. Senate speeches
Authors: Jan Vavra - PLUS (Austria) [presenting]
Bettina Gruen - Wirtschaftsuniversität Wien (Austria)
Paul Hofmarcher - University Salzburg (Austria)
Abstract: The world is evolving, and so is the vocabulary used to discuss various topics. However, current dynamic Poisson factorization models are designed for repeated observations on the same units, yielding matrices of counts for each considered time period. Such a format is not suitable for text-modelling purposes as documents usually cover very few topics, and their aggregation complicates the identifiability of the topics. Therefore, a time-varying Poisson factorization (TVPF) model is proposed that works with documents as the smallest observable units and captures the evolution of the popularity of words for each topic separately. The posterior distribution is approximated with variational inference, where the use of the mean-field variational family is questioned, especially for the time-varying components. Moreover, the frequently used random walk scheme is relaxed to a general AR(1) process. The stability and similarity of the topics are explored via a dissimilarity measure derived from the Kullback-Leibler divergence between variational families. The use of TVPF is illustrated in speeches from 18 sessions in the U.S. Senate, 1981-2016, where the primary focus is on the evolution of the climate change topic. Yet, for this example, both provided variational information criteria, and the simplicity of random walk parameterization estimated by mean-field variational inference is still preferred.
