Abstract
Topic models provide a useful method for inf dimensionality reduction and exploratory Ste data analysis in large text corpora. Most apRec proaches to topic model learning have been on based on a maximum likelihood objective. for Efficient algorithms exist that attempt to pro approximate this objective, but they have no dat provable guarantees. Recently, algorithms mod have been introduced that provide provable goa bounds, but these algorithms are not practitim cal because they are inefficient and not robust to violations of model assumptions. In this Aro paper we present an algorithm for learning abl topic models that is both provable and practha tical. The algorithm produces results comhas parable to the best MCMC implementations men while running orders of magnitude faster. tha ics alg