Performance evaluation of SVM and iterative FSVM classifiers with bootstrapping-based over-sampling and under-sampling

With the aim of a better generalizing performance than standard SVMs by assigning a fuzzy membership degree to each input point in the classification problem, Fuzzy Support Vector Machines (FSVMs) have been widely applied and have shown great success in various applications. However, the success of such a machine learning technique is inherently limited when applied to the problem of class imbalance learning as their structure is designed to generalize from sampled data sets. In this Paper, a new iterative fuzzy support vector machine algorithm (IFSVM) is proposed for severely imbalanced rail data classification with bootstrapping-based oversampling and undersampling. Data resampling techniques, oversampling and undersampling, are the best choice for overcoming the class imbalance problem. In this work, we succeeded in incorporating the unique learning mechanism of IFSVM and the class distribution advantages of resampling techniques. Experimental results on rail data set are effective not only on the overall generalization performance, but also on drastically reducing the algorithm´s time complexity and the number of support vectors. Our comparison study shows that the influence of data under-sampling method is superior compared to the bootstrapping-based over-sampling as it enables the proposed Iterative Fuzzy Support Vector machine to include better generalization capabilities with computational effectiveness.