Fast Detection of Overlapping Communities via Online Tensor Methods

Abstract

We present a fast tensor-based approach for detecting hidden overlapping communities under the Mixed Membership Stochastic Blockmodel (MMSB). We present two implementations, viz., a GPU-based implementation which exploits the parallelism of SIMD architectures and a CPU-based implementation for larger datasets, wherein the GPU memory does not suffice. Our GPU-based implementation involves a careful optimization of storage, data transfer and matrix computations. Our CPU-based implementation involves sparse linear algebraic operations which exploit the data sparsity. We use stochastic gradient descent for multilinear spectral optimization and this allows for flexibility in the tradeoff between node sub-sampling and accuracy of the results. We validate our results on datasets from Facebook, Yelp and DBLP where ground truth is available, using notions of p-values and false discovery rates, and obtain high accuracy for membership recovery. We compare our results, both in terms of execution time and accuracy, to the state-of-the-art algorithms such as the variational method, and report many orders of magnitude gain in the execution time. The tensor method is also applicable for unsupervised learning of a wide range of latent variable models, and we also demonstrate efficient recovery of topics from the New York Times dataset.