Semiparametric Subsampling and Data Condensation for Large-Scale Data Analytics

Subsampling is often used to reduce the complexity of large datasets. However, such methods need to ensure that the subsampled data are representative of the original dataset. Here, we introduce a new clustering-based data condensation (subsampling) framework for large datasets. The framework relies on the use of stratified sampling, Voronoi diagrams, and variational Bayes-based Gaussian mixture clustering. We tested the proposed framework on three large imbalanced benchmark datasets, namely cod-RNA, ds1.10, and ds1.100. The efficiency and generality of the proposed framework were assessed by comparing the predictive performance of the reduced datasets with the original datasets over two machine-learning classifiers, namely the random forest, and the radial basis function network. The evaluation metrics included the accuracy, F-measure and reduction percentage. We found that very high reduction percentages can be achieved using our new framework while maintaining satisfactory predictive performance.