Application of the Modified Imputation Method to Missing Data to Increase Classification Performance

Incomplete data or missing data diminishes the effectivity of statistical results, and may cause bias estimates, which in turn leads to unsound judgment. Inefficiency and impediments in data treatment analysis, which are among the predicaments linked with missing values, may affect the supervised learning process and reduce the classification accuracy and performance of the prediction model in a data mining task. This study applied the modified imputation method–which was previously tested with well-known imputation algorithms–to renowned classification techniques namely Naive Bayes, One-R, k-Nearest Neighbor (kNN), C4.5, and Support Vector Machine (SVM) using open data sets from the UCI Repository. The level of performance in terms of precision, accuracy, and Receiver Operating Characteristics (ROC) using Weka tool, before and after imputation was examined. This study manifests that there was an improvement in the classification performance upon the application of the modified imputation method on datasets during preprocessing, compared to that of datasets with missing values.