Fine-Tune: Integrating Fairness Constraints into Automated Hyperparameter Optimization for Bias-Resistant ML Models

One of the most challenging tasks for Machine Learning (ML) researchers is designing or implementing a decision model that is both fair and bias-free. The rapid adoption of Artificial Intelligence(AI) and ML technologies in the decision-making process has a significant impact and raises concerns regarding the fairness and bias of the model. The bias in ML models leads to discriminatory outcomes and reinforces societal inequalities. Unfair and biased models have a high impact on the accuracy of the model, and especially in the case of healthcare prediction, the model must be less biased and more fair. This research paper presents a comprehensive study of the diabetes dataset and explores strategies to mitigate the bias of the ML model. In the first phase of implementation, the ML models were trained on a biased dataset, and in the second phase, the models were trained on unbiased data. We deployed the SMOTE function to achieve fair and unbiased data. The K-Nearest Neighbor, Decision Tree, Random Forest, and Linear Regression ML models are used for the classification of the binary diabetes dataset. The accuracy of all four ML classifiers was evaluated on both biased and unbiased datasets. The accuracy on the biased dataset was reported as a higher score compared to the unbiased dataset due to the overfitting problem.