One-class ensemble classifier for data imbalance problems

Imbalanced data classification is an important issue in machine learning. Despite various studies, solving the data imbalance problem is still difficult. Since the oversampling method uses fake minority data, such a method is untrusted and causing security instability. The main objective of this paper is to improve accuracy for data imbalance classification without generating fake minority data. For this purpose, a reliable strategy is proposed using an ensemble of one-class classifiers. Such a classifier does not suffer data imbalance problems since the model learns from a single class. In particular, training data is split into minority and majority sets. Then, one-class classifiers are trained separately and applied to compute minority and majority scores for testing data. Finally, classification is made based on the combination of both scores. The proposed method is experimented with using imbalanced-learn datasets. Moreover, the result is compared with sampling methods via Decision Tree and K Nearest Neighbors classifiers. One-class ensemble classifier outperforms sampling methods in 20 datasets.

Automated summary

This paper addresses the issue of imbalanced data classification by proposing a reliable strategy using an ensemble of one-class classifiers, which avoids the trust issues and security instabilities caused by oversampling methods. Experimental results show that the one-class ensemble classifier outperforms sampling methods in 20 datasets.