Researchers Annalisa Deiana and Abdulhakim A. Qahtan have proposed a preprocessing framework designed to mitigate representation bias in machine learning models. Published in the journal Applied Intelligence, the study addresses the growing concern that algorithmic decision-making systems can perpetuate discrimination against individuals or groups identified by sensitive attributes such as gender or race. The authors argue that this bias often stems from imbalanced representation in training data, specifically in the distribution of favorable class labels across demographic groups. Their approach focuses on the training phase, aiming to create a balanced and fair dataset before the machine learning algorithms are trained.
The core of the proposed method involves the generation of synthetic data for minority groups. While synthetic data generation has been recognized as an effective technique for addressing bias, existing methods often fail to accurately identify the specific regions in the data space where adding synthetic examples is most effective. The new framework seeks to overcome this limitation by first identifying subgroups that have lower ratios of receiving favorable class labels. It then identifies the boundaries of the group examples with favorable class labels.
Once the boundaries are identified, the framework utilizes SMOTE-NC, a known oversampling technique, to generate synthetic examples around the group boundary. This process continues until the group reaches a target imbalance ratio. The authors emphasize that this synthetic data generation is performed carefully to minimize any negative impact on the overall performance of the machine learning models.
The study evaluates the performance and fairness of the proposed framework using standard performance and fairness measures. The results indicate that the framework successfully reduces representation bias and enhances fairness across demographic groups. Importantly, the authors note that this improvement in fairness is achieved while maintaining competitive predictive performance relative to state-of-the-art bias mitigation techniques.
Subject of Research: A preprocessing framework for generating synthetic data to mitigate representation bias in machine learning models.
Article Title: An oversampling framework for mitigating representation bias
Article References: Deiana, A., & Qahtan, A. A. (2026). An oversampling framework for mitigating representation bias. Applied Intelligence, 56(16), Article 547. https://doi.org/10.1007/s10489-026-07551-0
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07551-0
Keywords: machine learning, representation bias, synthetic data, fairness, oversampling, SMOTE-NC, data preprocessing, algorithmic bias, Applied Intelligence, framework, mitigating, representation
News Source: Blake Davidson. (November 8, 2026). New Framework Uses Synthetic Data to Reduce ML Bias. Scienmag.
