A Genetic Programming-Based Imputation Method for Classification with Missing Data

Many industrial and real-world datasets suffer from an unavoidable problem of missing values. The ability to deal with missing values is an essential requirement for classification because inadequate treatment of missing values may lead to large errors on classification. The problem of missing data has been addressed extensively in the statistics literature, and also, but to a lesser extent in the classification literature. One of the most popular approaches to deal with missing data is to use imputation methods to fill missing values with plausible values. Some powerful imputation methods such as regression-based imputations in MICE [36] are often suitable for batch imputation tasks. However, they are often expensive to impute missing values for every single incomplete instance in the unseen set for classification. This paper proposes a genetic programming-based imputation (GPI) method for classification with missing data that uses genetic programming as a regression method to impute missing values. The experiments on six benchmark datasets and five popular classifiers compare GPI with five other popular and advanced regression-based imputation methods in MICE on two measures: classification accuracy and computation time. The results showed that, in most cases, GPI achieves classification accuracy at least as good as the other imputation methods, and sometimes significantly better. However, using GPI to impute missing values for every single incomplete instance is dramatically faster than the other imputation methods.