Integrated approaches for improving the performance of imbalanced data classification

Abstract

The data generated by a number of real life applications are newlineimbalanced in nature. Class imbalance is a problem in which the class newlinedistribution is skewed. Hence the number of instances is not equal in each newlineclass label. In the disease diagnosis scenario, few people are only affected by newlinethe disease and the majority of people are healthy. The class label with the newlinemost number of instances is called majority class and the class label with a newlinelesser number of instances is called minority class. This problem arises in newlinemany applications such as natural disaster prediction, fraudulent transaction newlinedetection, and medical diagnosis. newlineThe class imbalance problem can be categorized into inter-class newlineand intra-class imbalance. The imbalance that exists in between the classes is newlinecalled inter-class or between-class imbalance. A single class containing a newlinenumber of sub clusters with unequal number of instances is called intra-class newlineimbalance or within-class imbalance. Naturally, the applications have either newlineinter-class imbalance or intra-class imbalance. But some applications have newlineboth inter-class imbalance and intra class imbalance. The class imbalance of newlineany data set is measured by class imbalance ratio, which is the ratio of the newlinenumber of minority class instances to majority class instances. The class newlineimbalance is a challenging aspect of building the classifier. newlineThere are various approaches developed to treat this imbalance newlineissue. The approaches are classified into four categories, namely data level, newlinealgorithm level, ensemble, and hybrid approaches. Data level methods newlineachieve class balancing by using sampling techniques. Some of the data level newlinemethods are Oversampling, Under-sampling and Hybrid sampling. newline