Large Scale Distributed Semi-Supervised Learning Using Streaming Approximation

Traditional graph-based semi-supervised learning (SSL) approaches, even widely applied, are not suited for massive data and large label since they scale linearly with the number of edges |E| and distinct m. To deal with the large label size problem, recent works propose-based methods to approximate the distribution on labels per node thereby a space reduction from O (m) to O (\ m), under certain. In this paper, we present a novel streaming graph-based SSL that captures the sparsity of the label distribution and ensures algorithm propagates labels accurately, and further reduces the space per node to O (1). We also provide a distributed version of the that scales well to large data sizes. Experiments on real-world demonstrate that the new method achieves better performance than state-of-the-art algorithms with significant reduction in memory. We also study different graph construction mechanisms for natural applications and propose a robust graph augmentation strategy trained state-of-the-art unsupervised deep learning architectures that yields significant quality gains.