Bagging and Feature Selection for Classification with Incomplete Data

Missing values are an unavoidable issue of many real-world datasets. Dealing with missing values is an essential requirement in classification problem, because inadequate treatment with missing values often leads to large classification errors. Some classifiers can directly work with incomplete data, but they often result in big classification errors and generate complex models. Feature selection and bagging have been successfully used to improve classification, but they are mainly applied to complete data. This paper proposes a combination of bagging and feature selection to improve classification with incomplete data. To achieve this purpose, a wrapper-based feature selection which can directly work with incomplete data is used to select suitable feature subsets for bagging. The experiments on eight incomplete datasets were designed to compare the proposed method with three other popular methods that are able to deal with incomplete data using C4.5/REPTree as classifiers and using Particle Swam Optimisation as a search technique in feature selection. Results show that the combination of bagging and feature selection can not only achieve better classification accuracy than the other methods but also generate less complex models compared to the bagging method.