Multiple Kernel Learning with Minority Oversampling for Classifying Imbalanced Data

Abstract

Class imbalance problems, developed due to the sampling bias or measurement error, occur frequently in real-world pattern classification tasks. The traditional classifiers focus on the overall classification accuracy and ignore the minority class, which may degrade the classification performance. However, existing oversampling algorithms generally make specific assumptions to balance the class size and do not sufficiently consider irregularities present in imbalanced data. As a result, these methods can perform well only on certain benchmarks. In this paper, by incorporating minority oversampling and cost-sensitive learning, we propose multiple kernel learning with minority oversampling (MKLMO), for efficiently handling the class imbalance problem with small disjuncts, overlapping, and nonlinear shape. Unlike existing methods where oversampling of the minority class is performed first and then a standard classifier is deployed on the rebalanced data, the proposed MKLMO generates synthetic instances and trains classifier synchronously in the same feature space. Specially, we define a distance metric in the optimal feature space by multiple kernel learning and use kernel trick to expand the original Gram matrix. Moreover, we assign different weights to instances, based on the imbalance ratio, for reducing the bias of the classifier towards the majority class. In order to evaluate the proposed MKLMO method, several experiments are performed with nine artificial and twenty-one real-world datasets. The experimental results show that our algorithm outperforms other baseline algorithms significantly in terms of the assessment metric geometric mean (G-mean), especially in the presence of data irregularities.