When Should We Ignore Examples with Missing Values?

Abstract

In practice, the dataset collected from data mining usually contains some missing values. It is common practice to perform case deletion by ignoring those data with missing values if the missing rate is certainly small. The aim of this paper is to answer the following question: When should one directly ignore sampled data with missing values? By using different types of datasets having various numbers of attributes, data samples, and classes, it is found that there are some specific patterns that can be considered for case deletion over different datasets without significant performance degradation. In particular, these patterns are extracted to act as the decision rules by a decision tree model. In addition, a comparison is made between cases with deletion and imputation over different datasets with the allowed missing rates and the decision rules. The results show that the classification performance results obtained by case deletion and imputation are similar, which demonstrates the reliability of the extracted decision rules.