RADAR: Rare Category Detection via Computation of Boundary Degree

Abstract

Rare category detection is an open challenge for active learning. It can help select anomalies and then query their class labels with human experts. Compared with traditional anomaly detection, this task does not focus on finding individual and isolated instances. Instead, it selects interesting and useful anomalies from small compact clusters. Furthermore, the goal of rare category detection is to request as few queries as possible to find at least one representative data point from each rare class. Previous research works can be divided into three major groups, model-based, density-based and clustering-based methods. Performance of these approaches is affected by the local densities of the rare classes. In this paper, we develop a density insensitive method for rare category detection called RADAR. It makes use of reverse k-nearest neighbors to measure the boundary degree of each data point, and then selects examples with high boundary degree for the class-label querying. Experimental results on both synthetic and real-world data sets demonstrate the effectiveness of our algorithm.