Self-regularized Lasso for selection of most informative features in microarray cancer classification

In this article, a new method is employed for maximizing the performance of the Least Absolute Shrinkage and Selection Operator (Lasso) feature selection model. In fact, we presented a novel regularization for the Lasso by employing an approach to find the best regularization parameter automatically which guarantees best performance of the Lasso in DNA microarray data classification. In our experiment, four well-known publicly available microarray datasets including breast cancer, Diffuse Large B-cell Lymphoma (DLBCL), leukemia and prostate cancer were utilized for evaluation the proposed methods. Experimental results demonstrated the significant dominance of the proposed Lasso against other widely used feature selection methods in terms of best features that led to best performance, robustness and stability in microarray data classification. Accordingly, the proposed method is a powerful algorithm for selection of most informative features which can be used for cancer diagnosis by gene expression profiles.