Machine learning for balanced and imbalanced data

Abstract

Recently, a kernel based online sequential extreme learning machine (OS-ELM) methods, OS-ELM with kernels (OS-ELMK) for non-stationary time series prediction is proposed. However, OS-ELMK has not been applied to classification problems and it is not clear which OS-ELM based method is more effective as a classifier. In this project, OS-ELM is extended to classification problems with relatively balanced datasets and compared with other OS-ELM methods. It is the first kernel-based OS-ELM classifier which can learn in both chunk-by-chunk and one-by-one modes. Guidelines for selecting appropriate OS-ELM classifier for different applications are also provided. Moreover, by combining OS-ELMK’s implicit feature mapping and a cost sensitive weighting scheme from weighted OS-ELM (WOS-ELM), a new kernel based online sequential method is proposed for imbalanced data classification. The new method is referred to as weighted OS-ELM with kernels (WOS-ELMK). The performance of WOS-ELMK is evaluated on benchmark imbalanced datasets and compared with a recently proposed voting based WOS-ELM (VWOS-ELM) method.