Borderline over-sampling for imbalanced data classification

Abstract

Abstract-Traditional classification algorithms, in many times, perform poorly on imbalanced data sets in which some classes are heavily outnumbered by the remaining classes. For this kind of data, minority class instances, which are usually much more of interest, are often misclassified. The paper proposes a method to deal with them by changing class distribution through oversampling at the borderline between the minority class and the majority class of the data set. A Support Vector Machines (SVMs) classifier then is trained to predict new unknown instances. Compared to other over-sampling methods, the proposed method focuses only on the minority class instances lying around the borderline due to the fact that this area is most crucial for establishing the decision boundary. Furthermore, new instances will be generated in such a manner that minority class area will be expanded further toward the side of the majority class at the places where there appear few majority class instances. Experimental results show that the proposed method can achieve better performance than some other over-sampling methods, especially with data sets having low degree of overlap due to its ability of expanding minority class area in such cases.

A comparative study on sampling techniques for handling class imbalance in streaming data

“…A more complex over-sampling technique interpolates synthetic minority instances between two existing ones [3]. Some studies have found that over-sampling of the minority class in borderline regions can provide better results [6], [10]. The simplest under-sampling technique is to randomly remove a number of majority instances, while a more intelligent technique [16] discards only those majority instances that are redundant, borderline, or noisy.…”

“…Standard learning algorithms usually make a bias toward the majority class to increase overall accuracy, and therefore reducing the predictive accuracy on the minority class. The most popular approach for handling the class imbalance problem is to rebalance the training set using sampling techniques [1]- [10]. An advantage of sampling is that we can simply apply a standard learning algorithm to a rebalanced training set, without a need to modify that algorithm.…”

PSO-Based Method for SVM Classification on Skewed Data-Sets

“…SMOTE introduces artificial instances in data sets by interpolating features values based on neighbors. In several studies have been shown that SMOTE is better than under-sampling and over-sampling techniques [3][4][5][6][7]. Moreover, SMOTE not cause any information loss and could potentially find hidden minority regions, because SMOTE identify similar but more specific regions in the feature space as the decision region for the minority class.…”

Oversampling the Minority Class in the Feature Space

“…However, the classes in general cannot be assumed to be convex, and hence the SMOTE does not avoid synthetic patterns to fall inside majority regions, therefore, more careful techniques have been developed to prevent this issue (prevent, but not solve). Adaptive synthetic [5]- [7] and cluster-based sampling methods [8], [9] are examples of more powerful techniques, based on extracting knowledge from the data to analyze which patterns and regions of the space are more suitable for oversampling. This will be referred in this paper to as preferential oversampling.…”

“…Ideally, a better fitted kernel will increase the class separability, providing a safer environment for the generation of synthetic patterns. The last part of this paper proposes a unified adaptive framework for preferential oversampling generalizing several oversampling approaches in the literature [3], [5], [6]. The optimal SVM hyperplane and kernel learning techniques are used for optimizing the synthetically generated patterns.…”