Gene selection for cancer classification using improved group lasso

An improved group lasso is proposed for simultaneous cancer classification and gene selection. A new criterion is firstly proposed to evaluate the individual gene importance by using the conditional mutual information. Then the weights with biological explanation are constructed and the improved group lasso is presented. A blockwise descent algorithm for solving the proposed model is also developed. The experimental results on lung cancer and prostate cancer data sets demonstrate that the proposed method can effectively perform classification and gene selection.