Sub-Gibbs Sampling: A New Strategy for Inferring LDA

Latent Dirichlet Allocation (LDA) has been widely used in text mining to discover topics from documents. One major approach to learn LDA is Gibbs sampling. The basic Collapsed Gibbs Sampling (CGS) algorithm requires O(NZ) computations to learn an LDA model with Z topics from a corpus containing N tokens. Existing approaches that improve the complexity of CGS focus on reducing the factor Z. In this work, we propose a novel and general Sub-Gibbs Sampling (SGS) strategy to improve the Gibbs-Sampling computation by reducing the sample space. This new strategy targets at reducing the factor N by sampling only a subset of the whole corpus. The design of the SGS strategy is based on two properties that we observe: (i) topic distributions of tokens are skewed and (ii) a subset of documents can approximately represent the semantics of the whole corpus. We prove that the SGS strategy can achieve comparable effectiveness (with bounded errors) and significantly reduce the complexity of existing Gibbs sampling algorithms. Extensive experiments on large real-world data sets show that the proposed SGS strategy is much faster than several state-of-the-art fast Gibbs sampling algorithms and the proposed SGS strategy can learn comparable LDA models as other Gibbs sampling algorithms.