Sampling based hybrid algorithms for imbalanced data classification

The microarray technology can exhibit the expression levels of tens of thousands of genes simultaneously, which helps to diagnose diseases particularly cancer at molecular level. But one of the most challenging issues associated with this technology is the skewed nature of the datasets, which makes the traditional classifiers inefficient in producing accurate classification results. However, a lot of work addressing this issue on binary class problems has been done by many researchers. This paper has combined three different sampling techniques namely, over sampling; under sampling and SMOTE with a meta-learning algorithm `DECORATE' to deal with a highly imbalanced multi-class microarray cancer dataset. The rate of accuracy of classification of the predictive models in case of imbalanced problem cannot be considered as an appropriate measure of effectiveness. Hence, different metrics are applied here to measure the performance of the proposed hybrid methods of classification. The experimental results show that unlike other traditional classification algorithms, our proposed hybrid methods are not sensitive to highly skewed multi-class microarray dataset.