Pick Your Neighborhood - Improving Labels and Neighborhood Structure for Label Propagation

Graph-based methods are very popular in semi-supervised learning due to their well founded theoretical background, intuitive interpretation of local neighborhood structure, and strong performance on a wide range of challenging learning problems. However, the success of these methods is highly dependent on the pre-existing neighborhood structure in the data used to construct the graph. In this paper, we use metric learning to improve this critical step by increasing the precision of the nearest neighbors and building our graph in this new metric space. We show that learning of neighborhood relations before constructing the graph consistently improves performance of two label propagation schemes on three different datasets - achieving the best performance reported on Caltech 101 to date. Furthermore, we question the predominant random draw of labels and advocate the importance of the choice of labeled examples. Orthogonal to active learning schemes, we investigate how domain knowledge can substantially increase performance in these semi-supervised learning settings.