| Imbalanced data problem raised great attentions in machine learning and data mining fields in recent decades. However, few efforts have been made on high dimensional feature selection of imbalanced dataset. This paper reviews existing methods, also raise two new methods for feature selection, one is the class decomposition method, and one is Hellinger distance method.In practical applications, the datasets are often imbalanced, the positive samples are rare, also the misclassification or omitting of these rare samples will lead to serious cost. Hence the research on classification of imbalanced data is economically meaningful. In chapter one we introduce some examples and briefly review the existing methods for imbalanced data. Methods could be classified into the five methods, they’re sampling method, data segmentation, cost-sensitive method, classifier combination method, and feature selection method. Chapter two reviews the first four methods, Chapter three reviews the feature selection method, including the Case-Specific-IG method, RELIEF method, FAST method and one framework for feature selection. In Chapter four, we raise the two new feature selection methods. The first method, the class decomposition method cuts classes into small pseudo classes using clustting algorithms and then using traditional feature selection measures to select those features which is significant for distinguishing classes. The second method uses Hellinger distance as measure for selection. Hellinger measures the distance between two distributions, hence the imbalance between two classes will not change the measure between the distributions of the two classes, hence Hellinger distance is skew-insensitive measure. The classification performance of the two raised methods applied to dataset of former years KDDCup are better compared with existing feature selection methods. |
