Distribution-Matched Imputation for Incomplete Data Evidential Classification

Abstract

Classifying missing data is an important and challenging topic in machine learning. However, the distribution between training and test sets may be inconsistent due to missing values, resulting in a negative impact on classification. To address this issue, we propose a novel distribution-matched imputation (DMI) method for classifying incomplete data based on evidential reasoning. Specifically, we consider the inconsistency in distribution between the training and test sets as the optimization objective to obtain optimal weights. Neighbors with different optimal weights are employed to estimate missing values, reducing the negative impact of inconsistent distribution on classification results. Then, we design a subspace-based evidential classification strategy to classify missing data with estimations, where the reliability of subclassification results consists of external and internal inconsistency. Doing this can characterize imprecision caused by inaccurate estimations and improve the classification performance of missing data. Our comprehensive experiments with various incomplete datasets reveal that the proposed DMI method offers more consistent and effective results compared to other related methods.