Multi-Objective Optimization in Machine Learning Model Training Using Evolutionary Algorithms and Pareto Front Analysis

Main Article Content

Swathy vodithala
Md. Sharfuddin waseem
Y. Bhavani
Kumar Dorthi
B. Surya Samantha
Nagaraju Krishna Chythanya

Abstract

In this study, we investigate multi-objective optimization (MOO) of machine learning models for a medical diagnosis task. We focus on predicting heart disease using the UCI Heart Disease dataset. We consider conflicting objectives such as maximizing accuracy, recall, and AUC while minimizing model complexity and ensuring fairness. To this end, we train and compare several hybrid models (e.g., CNN feature extractor + XGBoost classifier) and ensembles (e.g., SVM + MLP) under a multi-objective training paradigm. We apply evolutionary algorithms—NSGA-II, MOEA/D, and SPEA2—to optimize hyperparameters and model architectures over these objectives. The resulting Pareto fronts are analyzed to select optimal trade-offs. Experimental results include ROC curves, confusion matrices, loss curves, and Pareto plots. Our findings demonstrate that evolutionary multi-objective optimization can effectively balance performance and complexity in medical prediction tasks. We conclude with discussions of the accuracy–fairness trade-offs and outline future directions for multi-objective ML in healthcare.

Article Details

Section
Articles