Probabilistic topic modeling with latent Dirichlet allocation on Apache Spark

Abstract

In a world in which we have access to vast amount of data, it is important to develop new tools that allow us to navigate through it. Probabilistic topic models are statistical methods to analyse text corpora and discover themes that best explain its documents. In this work, we introduce probabilistic topic models with special focus on one of the most common models called Latent Dirichlet Allocation (LDA). To learn LDA model from data, we present two variational inference algorithms for batch and online learning. Both algorithms are implemented on a popular Big Data computing framework known as Apache Spark. We introduce this framework and We study the algorithm scalability and topic coherence in two different news data sets from New York Times and BBC News. The results point out to the need to tune up Apache Spark in order to boost its performance and to the goodness of the resulting topics in the BBC News dataset.