A Novel Assurance Procedure for Fair Data Augmentation in Machine Learning

In addressing the limited availability of data for predictive purposes with machine learning, we are concerned with potential biases arising from dataset augmentation. Despite advanced algorithms to generate synthetic data that can preserve the original data distribution, challenges remain, including the risk of perpetuating social biases. Our approach uses a similarity network representation that treats each data point as a node and strategically generates synthetic points near it. A vector label propagation algorithm, complemented by an exponential kernel for adjusting link weights, accurately labels these synthetic points. The primary goal is to reduce the system’s dependence on sensitive features without excluding them, thereby avoiding the risk of exacerbating biases or reducing data variation. Implemented in a big data ecosystem, our methodology enables continuous evaluation in an evolving domain, effectively addressing the challenges of data scarcity with a fairness-aware approach.