Online Inference of Topics with Latent Dirichlet Allocation
Kevin Canini, Lei Shi, Thomas Griffiths; JMLR W&CP 5:65-72, 2009.
Inference algorithms for topic models are typically designed to be run over an entire collection of documents after they have been observed. However, in many applications of these models, the collection grows over time, making it infeasible to run batch algorithms repeatedly. This problem can be addressed by using online algorithms, which update estimates of the topics as each document is observed. We introduce two related Rao-Blackwellized online inference algorithms for the latent Dirichlet allocation (LDA) model -- incremental Gibbs samplers and particle filters -- and compare their runtime and performance to that of existing algorithms.