Weight decision algorithm for oversampling technique on class-imbalanced learning
Youngil Kang, Sangchul Won · 2010
Oversampling technique is one of the methods to overcome the class imbalanced data problem by making new samples from existing one which belongs to minor class. In this paper, the weight decision algorithm for over-sampling minor samples in class-imbalanced learning is proposed. Weight decision algorithm determines the number of samples to populate from each sample aiming better classification performance than general over-sampling method. By applying edge detection algorithm to spatial space representation of training data, weights of minor samples are determined by calculating overall magnitude of gradient. The effect of weight decision algorithm is suggested by evaluating the classification results of over-sampled training data of several imbalanced datasets.