Improving performance for classification with incomplete data using wrapper-based feature selection

Cao Truong Tran, Mengjie Zhang, Peter Andreae, Bing Xue

Abstract: Missing values are an unavoidable problem of many real-world datasets. Inadequate treatment of missing values may result in large errors on classification; thus, dealing well with missing values is essential for classification. Feature selection has been well known for improving classification, but it has been seldom used for improving classification with incomplete datasets. Moreover, some classifiers such as C4.5 are able to directly classify incomplete datasets, but they often generate more complex classifiers with larger classification errors. The purpose of this paper is to propose a wrapper-based feature selection method to improve the ability of a classifier able to classify incomplete datasets. In order to achieve the purpose, the feature selection method evaluates feature subsets using a classifier able to classify incomplete datasets. Empirical results on 14 datasets using particle swarm optimisation for searching feature subsets and C4.5 for evaluating the feature subsets in the feature selection method show that the wrapper-based feature selection is not only able to improve classification accuracy of the classifier, but also able to reduce the size of trees generated by the classifier.