Abstract
In numerous real-world problems, we are faced with difficulties in learning from imbalanced data. The classification performance of a "standard" classifier (learning algorithm) is evidently hindered by the imbalanced distribution of data. The over-sampling and under-sampling methods have been researched extensively with the aim to increase the predication accuracy over the minority class. However, traditional under-sampling methods tend to ignore important characteristics pertinent to the majority class. In this paper, a novel under-sampling method based on information granules is proposed. The method exploits the concepts and algorithms of granular computing. First, information granules are built around the selected patterns coming from the majority class to capture the essence of the data belonging to this class. In the sequel, the resultant information granules are evaluated in terms of their quality and those with the highest specificity values are selected. Next, the selected numeric data are augmented by some weights implied by the size of information granules. Finally, a support vector machine and a K-nearest-neighbor classifier, both being regarded here as representative classifiers, are built based on the weighted data. Experimental studies are carried out using synthetic data as well as a suite of imbalanced data sets coming from the public machine learning repositories. The experimental results quantify the performance of support vector machine and K-nearest-neighbor with under-sampling method based on information granules. The results demonstrate the superiority of the performance obtained for these classifiers endowed with conventional under-sampling method. In general, the improvement of performance expressed in terms of G-means is over 10% when applying information granule under-sampling compared with random under-sampling.