Effect of Missing Data Types and Imputation Methods on Supervised Classifiers: An Evaluation Study

Data completeness is one of the most common challenges that hinder the performance of data analytics platforms. Different studies have assessed the effect of missing values on different classification models based on a single evaluation metric, namely, accuracy. However, accuracy on its own is a misleading measure of classifier performance because it does not consider unbalanced datasets. This paper presents an experimental study that assesses the effect of incomplete datasets on the performance of five classification models. The analysis was conducted with different ratios of missing values in six datasets that vary in size, type, and balance. Moreover, for unbiased analysis, the performance of the classifiers was measured using three different metrics, namely, the Matthews correlation coefficient (MCC), the F1-score, and accuracy. The results show that the sensitivity of the supervised classifiers to missing data differs according to a set of factors. The most significant factor is the missing data pattern and ratio, followed by the imputation method, and then the type, size, and balance of the dataset. The sensitivity of the classifiers when data are missing due to the Missing Completely At Random (MCAR) pattern is less than their sensitivity when data are missing due to the Missing Not At Random (MNAR) pattern. Furthermore, using the MCC as an evaluation measure better reflects the variation in the sensitivity of the classifiers to the missing data.