Alleviating Classification Problem of Imbalanced Dataset

The Class Imbalance problem occurs when there are many more instances of some class than others. i.e. skewed class distribution. In cases like this, standard classifier tends to be overwhelmed by the majority class and ignores the minority class. It is one of the 10 challenging problems of data mining research and pattern recognition. This imbalanced dataset degrades the performance of the classifier as accuracy is tendered towards the majority class. Several techniques have been proposed to solve this problem. This paper aims to improve the true positive rate/ detection of the minority class (GDM) which is the class of interest. This study proposes the use of two under sampling techniques reported in the literature. It involves under sampling the majority class which balances the dataset before classification. These under sampling schemes were evaluated on three learning algorithms (Decision tree both pruned and unpruned and RIPPER) using Matthew’s Correlation Coefficient (MCC) and Kappa Statistics as metrics. The implementation of these under sampling techniques was assessed in the medical domain. The real life dataset collected contained 886 instance of patients with diabetes mellitus disease. The diagnosis was in three classes with the following class distribution TYPE1 containing 62 instances, TYPE2 containing 807 instances and Gestation Diabetes Mellitus (GDM) which is the class of interest containing only 17 instances. This study revealed that, compared with the original dataset and RUS dataset, NCL dataset presents superiority in achieving better true positive rate for the minority class and also high MCC and KStatistics with the three learning algorithm.