Instance Selection for Imbalanced Data

Imbalanced data exhibit an unequal distribution with respect to the class labels. In a two-class imbalanced problem, elements of the majority class can vastly outnumber those belonging to the minority class. Several standard learning methods are hindered by such skewness present in the training set and fail to recognize minority instances in a posterior classification process. In real-world applications, e.g. in the medical domain or in the context of fraud detection, the minority class will usually be the class of interest. This motivates the development of techniques overcoming the challenges posed by data imbalance and ensuring an improvement of the classification performance. A considerable body of research (He & Garcia, 2009) has recently been done in this area. One prominent family of solutions are the resampling methods, which balance the dataset by introducing additional minority elements (oversampling), removing certain majority elements (undersampling) or a combination of both (hybrid methods).