Evolutionary isotonic separation for classification: theory and experiments

Abstract

Isotonic separation is a supervised machine learning technique where classification is represented as a linear programming problem (LPP) with an objective of minimizing the number of misclassifications. It is computationally expensive to solve the LPP using traditional methods when the dataset grows. Evolutionary isotonic separation (EIS), a hybrid classification algorithm, is introduced to tackle this issue. Here, isotonic separation acts as a host architecture where evolutionary framework based on genetic algorithm is embedded in the training phase of the isotonic separation, to find an optimum or near-optimum solution for the LPP. Evolutionary framework deploys a newly introduced slack vector to find the feasible solution. It also employs a position-based crossover operator to obtain the optimum or near-optimum solution. Experimental studies are conducted on Wisconsin Breast Cancer dataset and a synthetic dataset. Experimental and statistical results show that EIS outperforms its predecessors and state of the art machine learning techniques in terms of accuracy.