Online Belief Propagation for Topic Modeling

The batch latent Dirichlet allocation (LDA) algorithms play important roles in probabilistic topic modeling, but they are not suitable for processing big data streams due to high time and space compleixty. Online LDA algorithms can not only extract topics from big data streams with constant memory requirements, but also detect topic shifts as the data stream flows. In this paper, we present a novel and easy-to-implement online belief propagation (OBP) algorithm that infers the topic distribution from the previously unseen documents incrementally within the stochastic approximation framework. We discuss intrinsic relations between OBP and online expectation-maximization (OEM) algorithms, and show that OBP can converge to the local stationary point of the LDA's likelihood function. Extensive empirical studies confirm that OBP significantly reduces training time and memory usage while achieves a much lower predictive perplexity when compared with current state-of-the-art online LDA algorithms. Due to its ease of use, fast speed and low memory usage, OBP is a strong candidate for becoming the standard online LDA algorithm.