Far Point Algorithm: Active Semi-supervised Clustering for Rare Category Detection

Abstract

In some data sets the number of categories (i.e. classes) that are represented is not known in advance. The process of discovering these categories can be difficult, particularly when a data set is skewed, such that the number of data points of some classes may greatly exceed those of other classes. Rare category detection algorithms address this problem by trying to present a user with at least one data point from each category, while minimizing the overall number of data points presented. We present an algorithm based on active and semi-supervised learning that finds category clusters using a query selection strategy that maximizes the distance from a set of already labeled data points to a query data point. We evaluate the algorithm's performance on artificially skewed versions of the MNIST data set as a rare category detection algorithm, investigating differences in performance due to both the effects of relative frequency and inherent class structure differences in feature space.