Feature-Wise Mixing for Mitigating Contextual Bias in Predictive Supervised Learning

TL;DR

This paper proposes a feature-wise mixing framework to reduce bias in machine learning models by redistributing feature representations, achieving significant bias reduction and improved fairness without high computational costs.

Contribution

It introduces a novel feature-wise mixing method that mitigates contextual bias and outperforms existing bias mitigation techniques like SMOTE in efficiency and effectiveness.

Findings

Achieved an average bias reduction of 43.35%.

Significantly decreased mean squared error across classifiers.

Outperformed SMOTE oversampling in bias mitigation.

Abstract

Bias in predictive machine learning (ML) models is a fundamental challenge due to the skewed or unfair outcomes produced by biased models. Existing mitigation strategies rely on either post-hoc corrections or rigid constraints. However, emerging research claims that these techniques can limit scalability and reduce generalizability. To address this, this paper introduces a feature-wise mixing framework to mitigate contextual bias. This was done by redistributing feature representations across multiple contextual datasets. To assess feature-wise mixing's effectiveness, four ML classifiers were trained using cross-validation and evaluated with bias-sensitive loss functions, including disparity metrics and mean squared error (MSE), which served as a standard measure of predictive performance. The proposed method achieved an average bias reduction of 43.35% and a statistically significant decrease in MSE across all classifiers trained on mixed datasets. Additionally, benchmarking against established bias mitigation techniques found that feature-wise mixing consistently outperformed SMOTE oversampling and demonstrated competitive effectiveness without requiring explicit bias attribute identification. Feature-wise mixing efficiently avoids the computational overhead typically associated with fairness-aware learning algorithms. Future work could explore applying feature-wise mixing for real-world fields where accurate predictions are necessary.