Design-Based Subsampling for Labeling Large and High-Dimensional Datasets

Data labeling, the process of assigning labels or annotations to data points, is crucial in supervised machine learning for training models to make accurate predictions in various applications. Labels refer to the output variables that the machine learning model aims to predict or classify. For instance, in genetic and genomic studies, labels may refer to traits or the presence of diseases, and accurate data labeling is essential for training models to understand the relationships between genetic information and various traits or diseases. The labeling process is resource-intensive, requiring domain expertise, advanced experimentation techniques, and rigorous quality control to ensure accuracy. Consequently, labeling all data points in a large dataset is often impractical due to resource limitations. Therefore, selecting an informative subsample from the large pool of data points to label becomes a critical and challenging problem. This project aims to develop subsampling methods for labeling large and high-dimensional datasets. The anticipated results will be applicable to genetics, biology, and medicine. Graduate and undergraduate students will be involved in the project and exposed to these results, which will also be incorporated into university courses. The project aims to develop advanced subsampling techniques for data labeling, particularly for large and high-dimensional datasets. Optimal subsampling approaches for both continuous and binary labels will be developed to enhance the predictive performance of models trained on the labeled subsample. Additionally, sequential sampling plans will be investigated. The project also emphasizes fairness in the trained models, striving to develop subsampling methods that ensure a balanced representation of multiple demographic groups in the labeled data, achieving consistent accuracy across all demographic groups. This project will contribute substantially to advancing subsampling methodologies in statistics and data science. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.