Interpretable Components Using Genetic Programming Employing Instruction-like Structure

This paper introduces a novel feature extraction method, IGP, that generates components through both linear and non-linear combinations of features using Genetic Programming (GP). Unlike traditional GP approaches that rely on expression trees, IGP utilizes an instruction line structure. The study evaluates IGP’s performance against 5 established feature extraction methods across 23 datasets, encompassing binary and multiclass classification tasks. The results demonstrate that IGP excels in several instances, particularly in binary classification, with further analysis exploring how the relationship between the number of classes, features, and instances contributes to its performance. Additionally, the scope for future investigations of IGP are commented.