Fairness-Aware Learning Framework For Bias Mitigation In Algorithmic Decision Systems

Abstract

The increased reliance on Machine Learning (ML) based algorithmic decisions has given rise to some critical issues such as fairness and bias because the machine learning models, if trained on biased data, tend to make discriminatory predictions. Therefore, to counter these problems, this study puts forth an innovative Fairness-Aware Grouping Learning Framework (FAGLF). To improve the quality of the datasets, missing value imputation is utilized to replace incomplete data points, while the Synthetic Minority Over-sampling Technique (SMOTE) addresses the problem of class imbalance. Furthermore, Independent Component Analysis (ICA) based feature extraction is employed to enhance data quality and enable more equitable predictive learning. The approach embeds data instances into a feature space and organizes them into homogeneous subgroups based on similarity patterns. For each subgroup, a specialized predictive model is trained, allowing more precise learning while maintaining overall consistency. The framework is formulated as a three-level optimization problem, solved using Dynamic Transient Search Optimized Extreme Gradient Boosting (DTSO-XGBoost), which balances exploration and exploitation during optimization. For better performance on the smaller and class-imbalanced datasets, techniques for domain adaptation have been used. Fairness is ensured by keeping equal distribution among sensitive variables such as gender and socio-economic status. The developed model is tested on the dataset "Fairness Decision Records" using fairness metrics implemented in Python. It achieves a precision of 0.972, a recall of 0.948, and an F1 score of 0.960. Results demonstrate significant improvement in fairness–accuracy trade-offs while maintaining strong predictive performance, highlighting the effectiveness of the proposed framework.