Imputations for High Missing Rate Data in Covariates Via Semi-supervised Learning Approach

Advancements in data collection techniques and the heterogeneity of resources can yield high percentages missing observations on variables, such as block-wise data. Under missing-data scenarios, traditional methods simple average, k-nearest neighbor, multiple, regression imputations may lead to results that are unstable or unable be computed. Motivated by concept semi-supervised learning, we propose a novel approach with which fill values covariates have rates. Specifically, consider nonmissing subjects any covariate unlabeled labeled target outputs, respectively, treat their corresponding responses inputs. This innovative setting allows us impute large number without imposing model assumptions. In addition, resulting imputation has closed form for continuous covariates, it calculated efficiently. An analogous procedure is applicable discrete covariates. We further employ nonparametric show theoretical properties imputed Simulation studies an online consumer finance example presented illustrate usefulness proposed method.