Fuzzy multiclass support vector machines for unbalanced data
Yuanyuan Wu, Li‐Yong Shen, Sanguo Zhang · 2017
Traditional support vector machines have low classification performance on unbalanced datasets and are more influenced by dataset noise, which can lead to a deviation of the classification results. To address this problem, other studies have proposed fuzzy support vector machines for unbalanced datasets. However, most of these algorithms are directed to the second-class classification problem and their fuzzy membership functions typically only consider the distance factor, failing to accurately reflect the importance of training sample points. In this paper, a fuzzy multiclass support vector machine algorithm for unbalanced data is proposed. The algorithm uses the distance from the training sample point to the center of its class and a weighted class-overlap method to design the sample fuzzy membership function. Corresponding membership values are assigned according to the importance of the sample points: the weight of the support vector points is increased and the weight of noise is decreased. The latest unbalanced adjustment factors are used to reduce the influence of the unbalanced data on the classification results. Experimental results confirm that, compared with traditional fuzzy support vector machines, the proposed algorithm can address unbalanced data and noise more effectively in multi-classification problems.