Active learning with missing values considering imputation uncertainty

Abstract

To perform active learning with incomplete datasets containing missing values, conventional methods first impute the missing values via missing value imputation. Unlabeled instances are then sequentially selected to be labeled based on an acquisition function towards improving the prediction model. However, inaccurate imputations would negatively affect the performance of the prediction model. In this study, we propose a novel query selection method that considers the imputation uncertainty for active learning with missing values. We quantify the imputation uncertainty of each instance using multiple imputation. For query selection, the acquisition function is penalized by the imputation uncertainty. Consequently, inaccurately imputed instances are less likely to be selected for labeling, thereby avoiding superfluous labeling costs. We verified the effectiveness of the proposed method on twenty benchmark datasets using various missing rates and prediction models.