Balanced Subsampling for Big Data with Categorical Covariates
share
The use and analysis of massive data are challenging due to the high storage and computational cost. Subsampling algorithms are popular to downsize the data volume and reduce the computational burden. Existing subsampling approaches focus on data with numerical covariates. Although big data with categorical covariates are frequently encountered in many disciplines, the subsampling plan has not been well established. In this paper, we propose a balanced subsampling approach for reducing data with categorical covariates. The selected subsample achieves a combinatorial balance among values of covariates and therefore enjoys three desired merits. First, a balanced subsample is nonsingular and thus allows the estimation of all parameters in ANOVA regression. Second, it provides the optimal parameter estimation in the sense of minimizing the generalized variance of the estimated parameters. Third, the model trained on a balanced subsample provides robust predictions in the sense of minimizing the worst-case prediction error. We demonstrate the usefulness of the balanced subsampling over existing data reduction methods in extensive simulation studies and a real-world application.
READ FULL TEXT