A multi-objective pareto-optimal wrapper based framework for cancer-related gene selection

Abstract

. In this paper, a multi-criteria feature selection framework is proposed to integrate the wrapper method and Pareto Optimal (PO) method so that cancer-related genes in microarray datasets can be identified. Sequential forward selection is applied for feature selection among cross-validated training sets, and PO is employed as an aggregation method to combine wrapper-based gene selection results from the training sets. The proposed gene selection does not require user intervention and PO also selects each valuable gene when structuring the most representative gene subset. In order to test the performance of the proposed framework, an experimental study has been conducted on three publicly available cancer microarray datasets. The results show that the proposed framework gives robust aggregation and the accuracy is boosted when feature selection results are combined with PO. The findings also demonstrate that the Pareto Optimality based framework is robust against variations in the training sets and is less prone to over fitting.