MIAC: Mutual-Information Classifier with ADASYN for Imbalanced Classification

Abstract

currently, classification of imbalanced data is a significant issue in the area of data mining and machine learning because of the imbalance of most of the data set. An effective solution of this problem is Cost-Sensitive Learning (CSL), but when the costs are not given, this method cannot work property. As a Cost-Free Learning (CFL) method, Mutual-Information Classification (MIC) can obtain the optimal classification results when the cost information is not given. But this method emphasizes the data of minority class too much and neglects the accuracy of the classification of majority class. And based on the above, this paper presented a CFL method called Mutual-Information-ADASYN Classification (MIAC). Firstly, we get the abstaining samples which are hard to be classified by using MIC. Then we use these abstention samples to synthesize new instance by using the method of ADASYN. Thirdly, we build Mutual- Information-ADASYN Classification using the new samples. Finally, we use our classifier to get the final results. We evaluated the performance of MIAC on several imbalance binary datasets with different imbalance ratios. The experimental results indicate that the MIAC is more effective than MIC on dealing with imbalanced datasets.