Unified Strategy for Feature Selection and Data Imputation

Data-related issues represent the main causes for insufficient performance in data mining. Existing strategies for tackling these issues include procedures for handling incomplete data - mandatory in various schemes, and feature selection, both augmenting the learning process. Our previous work on data imputation has shown that a good imputation policy for strongly correlated attributes with the class can improve the learning accuracy. Moreover, feature selection also enhances the performance of an inducer. The focus of the paper is to validate the performance and stability of our combined methodology for pre-processing data. The novelty of the method resides in the combination of feature selection with data imputation, in order to obtain an improved version of the training set. The experimental results have shown that, when mining incomplete data, our combined pre-processing methodology boosts the accuracy of a classifier. Moreover, it is more successful than each of the individual steps it combines, feature selection and imputation, producing better or similar results.