Performance of SVM with Multiple Kernel Learning for Classification Tasks of Imbalanced Datasets

Support vector machine (SVM) is one of the most popular algorithms in machine learning and data mining. However, its reduced efficiency is usually observed for imbalanced datasets. To improve the performance of SVM for binary imbalanced datasets, a new scheme based on oversampling and the hybrid algorithm were introduced. Besides the use of a single kernel function, SVM was applied with multiple kernel learning (MKL). A weighted linear combination was defined based on the linear kernel function, radial basis function (RBF kernel), and sigmoid kernel function for MKL. By generating the synthetic samples in the minority class, searching the best choices of the SVM parameters and identifying the weights of MKL by minimizing the objective function, the improved performance of SVM was observed. To prove the strength of the proposed scheme, an experimental study, including noisy borderline and real imbalanced datasets was conducted. SVM was applied with linear kernel function, RBF kernel, sigmoid kernel function and MKL on all datasets. The performance of SVM with all kernel functions was evaluated by using sensitivity, G Mean, and F measure. A significantly improved performance of SVM with MKL was observed by applying the proposed scheme.