ZenLDA: Large-Scale Topic Model Training on Distributed Data-Parallel Platform

Recently, topic models such as Latent Dirichlet Allocation (LDA) have been widely used in large-scale web mining. Many large-scale LDA training systems have been developed, which usually prefer a customized design from top to bottom with sophisticated synchronization support. We propose an LDA training system named ZenLDA, which follows a generalized design for the distributed data-parallel platform. The novelty of ZenLDA consists of three main aspects: (1) it converts the commonly used serial Collapsed Gibbs Sampling (CGS) inference algorithm to a Monte-Carlo Collapsed Bayesian (MCCB) estimation method, which is embarrassingly parallel; (2) it decomposes the LDA inference formula into parts that can be sampled more efficiently to reduce computation complexity; (3) it proposes a distributed LDA training framework, which represents the corpus as a directed graph with the parameters annotated as corresponding vertices and implements ZenLDA and other well-known inference methods based on Spark. Experimental results indicate that MCCB converges with accuracy similar to that of CGS, while running much faster. On top of MCCB, the ZenLDA formula decomposition achieved the fastest speed among other well-known inference methods. ZenLDA also showed good scalability when dealing with large-scale topic models on the data-parallel platform. Overall, ZenLDA could achieve comparable and even better computing performance with state-of-the-art dedicated systems.