Semi-Supervised Learning with Competitive Infection Models

The goal of semi-supervised learning methods is to effectively combine labeled and unlabeled data to arrive at a better model. Many methods rely on graph-based approaches, where labels are propagated through a graph over the input examples. In most current methods, the propagation mechanism underlying the learning objective is based on random walks. While theoretically elegant, random walks suffer from several drawbacks which can hurt predictive performance. In this work, we explore dynamic infection processes as an alternative propagation mechanism. In these, unlabeled nodes can be "infected" with the label of their already infected neighbors. We provide an efficient, scalable, and parallelizable algorithm for estimating the expected infection outcomes. We also describe an optimization view of the method, relating it to Laplacian approaches. Finally, experiments demonstrate that the method is highly competitive across multiple benchmarks and for various learning settings.