Use of data mining for intelligent evaluation of imputation methods

Abstract

In real-world situations, researchers frequently face the difficulty of missing values (MV), i.e., values not observed in a data set. Data imputation techniques allow the estimation of MV using different algorithms, by means of which important data can be imputed for a particular instance. Most of the literature in this field deals with different imputation methods. However, few studies deal with a comparative evaluation of the different methods as to provide more appropriate guidelines for the selection of the method to be applied to impute data for specific situations. The objective of this work is to show a methodology for evaluating the performance of imputation methods by means of new metrics derived from data mining processes, using quality metrics of data mining models. We started from the complete dataset that was amputated with different amputation mechanisms to generate 63 datasets with MV; these were imputed using Median, k-NN, k-Means and Hot-Deck imputation methods. The performance of the imputation methods was evaluated using new metrics derived from quality metrics of the data mining processes, performed with the original full file and with the imputed files. This evaluation is not based on measuring the error when imputing (usual operation), but on considering the similarity of the values of the quality metrics of the data mining processes obtained with the original file and with the imputed files. The results show that –globally considered and according to the new proposed metric, the imputation methods that showed the best performance were k-NN and k-Means. An additional advantage of the proposed methodology is that it provides predictive data mining models that can be used a posteriori.