HAR-MI with COSTE in Handling Multi-Class Imbalance

The class imbalance problem is a serious problem in machine learning. This problem can occur in two-class and multi-class problems. This problem can result in low accuracy and not obtaining information regarding the minority class. Approaches to overcome this problem often use a combination of the Data-Level Approach and Algorithm-Level Approach, which is often referred to as a Hybrid Approach. One of the Hybrid Approach methods to solve the multi-class imbalance problem is the Hybrid Approach Redefinition-Multiclass Imbalance (HAR-MI). The sampling method used in the HAR-MI is the Oversampling method. Oversampling tends to be chosen because Undersampling can eliminate useful information, but Oversampling often causes Over-Fitting. Therefore, this study will modify the HAR-MI approach by using the Complexity-based OverSampling TEchnique (COSTE). COSTE will rank the instances based on the measurement of complexity so that it will provide better results and prevent Over-Fitting. The results showed that HAR-MI with COSTE gave better results than the classic HAR-MI.