Clustering-based binary-class classification for imbalanced data sets

Abstract

In this paper, we propose a new clustering-based binary-class classification framework that integrates the clustering technique into a binary-class classification approach to handle the imbalanced data sets. A binary-class classifier is designed to classify a set of data instances into two classes; while the clustering technique partitions the data instances into groups according to their similarity to each other. After applying a clustering algorithm, the data instances within the same group usually have a higher similarity, and the differences among the data instances between different groups should be larger. In our proposed framework, all negative data instances are first clustered into a set of negative groups. Next, the negative data instances in each negative group are combined with all positive data instances to construct a balanced binary-class data set. Finally, subspace models trained on these balanced binary-class data sets are integrated with the subspace model trained on the original imbalanced data set to form the proposed classification model. Experimental results demonstrate that our proposed classification framework performs better than the comparative classification approaches as well as the subspace modeling method trained on the original data set alone.