Investigation of the Impact of Missing Value Imputation Methods on the k-NN Classification Accuracy

摘要

Paper desribes results of an experiment where various scenarios of missing values occurrence in the data repository has been tested. Experiment was coducted on a publicly available database, containing complete, multidimensional continuous dataspace and multiple classes. Missing values were introduced using "completely at random" scheme. Tested scenarios were: training and testing using incomplete dataset, training on complete data set and testing on incomplete and vice versa. For comparison to data imputation methods also the ensemble of single-feature kNN classifiers, working withoud data imputation, has been tested.