A Machine Learning based Approach to Reduce Behavioral Noise Problem in an Imbalanced Data: Application to a fraud detection

Abstract

The question of class imbalance has become more pronounced with the application of learning algorithms in real applications. It has received significant attention in the machine learning and data mining community. This problem is present in fraud detection, medical diagnostics, and a number of other areas where training data contains significantly more representatives of one class (called the majority class) than the other class (called the minority class). Machine learning techniques struggle to deal with imbalanced data by focusing on minimizing the error rate for the majority class while ignoring the minority class, which is the most interesting from a learning point of view and also involves a high cost when it is not well classified. However, the imbalance ratio is not the only cause of poor performance when learning from imbalanced data. Another critical factor that accompanies imbalanced data in the real world is the presence of a number of instances of the two classes being overlapped in feature space. This problem is commonly referred to as class overlap and we have called it “behavioral noise”. In this paper, we propose One Side Behavioral Noise Reduction (OSBNR) approach to deal with the problem of class imbalance in the presence of a behavioral noise level. OSBNR is based on two stages. Firstly, a clustering is applied to groups similar instances of the minority class in multiple behavior clusters. Secondly, we select and eliminate instances of the majority class, considered as behavioral noise, which overlap with the behavior clusters of the minority class. The results of experiments conducted on a representative public dataset confirm that the proposed approach is effective for class imbalance problem in the presence of behavioral noise.