Fast and Robust Fuzzy-based Hybrid Data-level Method to Handle Class Imbalance

Abstract

Conventional classification algorithms do not provide accurate results when the data distribution (class sizes) is unequal or data is corrupted with noise because the results are biased towards the bigger class. In many real life cases, there is a requirement to uncover unusual/smaller classes. There are a bundle of examples where importance of smaller/rare class is much-much higher than the bigger class for example-brain tumor detection, credit card fraud or anomaly detection and many more. This is usually called as problem of imbalance classes. The situation becomes worst when the data is corrupted with extra impurities like noise in data or overlapping of class or any other glitch in data because in this scenario traditional methods produce more poor results. This paper proposed a fast, simple and effective data level hybrid technique based on fuzzy concept to overcome the class imbalance problem in noisy condition. To appraise the classification performance of the offered technique it is tested with 40 UCI real imbalanced data sets having imbalance ratio ranges from 1.82 to 129.44 and compared with 12 other approaches. The outcome specifies that the presented hybrid data level technique performed better and in a fast manner when compared to other approaches