Dealing with Class Imbalance the Scalable Way: Evaluation of Various Techniques Based on Classification Grade and Computational Complexity

Abstract

Highly imbalanced datasets continue to be a challenge in many data mining applications. It is surprising that state-of-the-art techniques countering class imbalances are usually very computationally expensive and therefore unscalable. Most research effort has been directed into enhancing those techniques, e.g., by focusing on borderline examples or combining multiple techniques. This is usually accompanied by an increased computational complexity, reducing the scalability even further. This article has four major contributions: First, existing techniques to deal with imbalanced datasets are evaluated regarding their computational cost and influence on classification performance on a variety of publicly available datasets and classifiers. Second, a new, scalable technique, class specific scaling (CSS) is proposed as an alternative and compared to the existing techniques. Third, a parameter free class overlap and noise measure is introduced to complement the existing measures to assess the dataset's properties, such as the class balance ratio, and the number of features and samples. This enables a finer categorization of imbalanced datasets. Fourth, based on these measures and basic conditions such as scalability and the used classifier, general recommendations regarding the suitability of the different techniques are derived.