Bias Detection and Fairness Optimization in Algorithmic Systems

Abstract

Algorithmic decision-making systems are widely deployed in critical domains, where embedded biases can lead to systematic discrimination against sensitive groups and reliability concerns. Existing fairness approaches are often stage-specific and fail to provide a unified mechanism that simultaneously detects bias and optimizes fairness without significantly compromising predictive performance. This research offers an integrated model for bias detection and fairness optimization in algorithmic systems to ensure equitable and robust decision-making. A fairness-aware structured dataset of 5,000 samples includes demographic-sensitive attributes, behavioral features, and decision-related variables. Preprocessing is performed using re-weighting and synthetic data balancing (SMOTE) to address class and group imbalance. Feature extraction employs Principal Component Analysis (PCA) to transform features into an uncorrelated space while preserving variance. The proposed Adaptive Gradient-Constrained Fair Logistic Regression (AGCFLR) is designed to minimize bias by enforcing fairness-aware constraints during model training while maintaining predictive performance. It integrates Adaptive Gradient-Based Constrained Optimization (AGCO) to dynamically adjust model parameters under fairness constraints and Logistic Regression (LR) to provide interpretable and efficient classification for decision-making tasks. The proposed AGCFLR model achieved superior performance with 94.8% accuracy, 93.1% precision, and 92.7% recall using Python 3.10. It effectively reduces bias across sensitive groups while preserving overall classification performance, demonstrating balanced trade-off accuracy. It demonstrates that fairness-aware optimization can effectively mitigate bias while maintaining model reliability.