This project aims to impelement a parallel version of the Latent Dirichlet Allocation algorithm which is one of the most used topic modeling algorithm in the industry. For theoretcial details, please take a look on the papers.
The core idea consists in running many concurrent Gibbs samplers. Each gibbs sampler compute its own documents-counts and words-counts matrices. At the end of each step, the words-counts matrix of each sampler is communicated to the master which is in charge of updating of the whole set of matrices. This is repeated until convergence ...
Please take check the pLda notebook for detailed, setp by step presentation of our impementation.