An empirical study on the feature selection abilities of SLIM-GSGP [poster]

Abstract

Feature Selection (FS) is a key characteristic of any Machine Learning method. Genetic Programming (GP) performs it inherently, using evolution pressure to exclude redundant or irrelevant features. However, this ability is lost in Geometric Semantic Genetic Programming (GSGP), where Geometric Semantic Operator (GSO) keep adding genetic material to the individuals, inevitably adding noisy features. This work focuses on comparing the FS abilities of GSGP and Semantic Learning algorithm based on Inflate and deflate Mutations (SLIM), a promising new variant that employs Deflate Geometric Semantic Mutation (DGSM), a genetic operator that is able to remove genetic material while still inducing an unimodal fitness landscape. The experimental results show how SLIM has superior FS abilities compared to GSGP.