Method for Classification of Cancers with Partial Least Squares Regression as Feature Selector with Kernel SVM

Classification of cancers according to the exact place of origin in the body is an important research problem that is being addressed both clinically and computationally. Application of data science and machine learning to a vast volume of imaging and genomics data regarding cancers has enabled computational researchers to accurately classify the tumor samples according to their place of origin. In current work, we developed a method to classify tumors using partial least squares regression as feature selector and support vector machine classifiers. We have evaluated our approach on three cancer gene expression datasets and found classification accuracies of 100% in some circumstances. The comparison is conducted with standard classification methods like Decision Trees and simple Support Vector Machines with respect to standard performance parameters and the time taken for classification. The comparison in terms of training and testing accuracies and the time taken for classification results show that our method performs consistently better than conventional methods.