Genetic Programming for Feature Subset Ranking in Binary Classification Problems

Kourosh Neshatian, Mengjie Zhang

Abstract: We propose a genetic programming (GP) system for measuring the relevance of subsets of features in binary classification tasks. A virtual program structure and an evaluation function are defined in a way that constructed GP programs can measure the goodness of subsets of features. The proposed system can detect relevant subsets of features in different situations including multimodal class distributions and mutually correlated features where other ranking methods have difficulties. Our empirical results indicate that the proposed system is good at ranking subsets and giving insight into the actual classification performance. The proposed ranking system is also efficient in terms of feature selection.