MODis: Generating Skyline Datasets for Data Science Models

Preparing high-quality datasets for data science models is a critical yet challenging task. Traditional data discovery approaches often focus on a single optimization criterion, which can introduce bias and limit downstream performance. This paper introduces MODis, a multi-objective data discovery framework that generates skyline datasets tailored to user-defined model performance measures (e.g., accuracy, training cost). MODis formalizes the skyline dataset generation problem using a finite state transducer (FST), and provides three algorithms:(1) a reduce-from-universal strategy that prunes unpromising data while approximating Pareto-optimal sets, (2) a bi-directional search that interleaves augmentation and reduction with correlation-based pruning, and (3) a diversification algorithm to mitigate dataset bias. Experiments with benchmark datasets and scientific applications show that MODis can improve model accuracy by up to 2× while reducing training cost, outperforming baseline data integration and feature selection methods.

In Proceedings of the 28th International Conference on Extending Database Technology (EDBT)