Multiple Imputation and Ensemble Learning for Classification with Incomplete Data

Missing values are a common issue in many real-world datasets, and therefore coping with such datasets is an essential requirement of classification since inadequate treatment of missing values often leads to large classification errors. One of the most popular ways to address incomplete data is to use imputation methods to fill missing fields with plausible values. Multiple imputation, which fills each missing field with a set of plausible values, is a powerful approach to dealing with incomplete data, but is mainly used for statistical analysis. Ensemble learning which constructs a set of classifiers instead of one classifier has proven capable of improving classification accuracy, but has been mainly applied to complete data. This paper proposes a combination of multiple imputation and ensemble learning to build an ensemble of classifiers for incomplete data classification tasks. A multiple imputation method is used to generate a set of diverse imputed datasets which is then used to build a set of diverse classifiers. Experiments on ten benchmark datasets use a decision tree as classification algorithm and compare the proposed approach with two other popular approaches to dealing with incomplete data. The results show that, in almost all cases, the proposed method achieves significantly better classification accuracy than the other methods.