A new preprocessing framework generates synthetic data to reduce representation bias in machine learning models while maintaining competitive predictive performance.