Data Cleansing Meets Feature Selection: A Supervised Machine Learning Approach

Abstract

This paper presents a novel procedure to apply in a sequential way two data preparation techniques from a different nature such as data cleansing and feature selection. For the former we have experienced with a partial removal of outliers via inter-quartile range whereas for the latter we have chosen relevant attributes with two widespread feature subset selectors like CFS (Correlation-based Feature Selection) and CNS (Consistency-based Feature Selection), which are founded on correlation and consistency measures, respectively. Empirical results on seven difficult binary and multi-class data sets, that is, with a test error rate of at least a 10%, according to accuracy, with C4.5 or 1-nearest neighbour classifiers without any kind of prior data pre-processing are outlined. Non-parametric statistical tests assert that the meeting of the aforementioned two data preparation strategies using a correlation measure for feature selection with C4.5 algorithm is significant better, measured with roc measure, than the single application of the data cleansing approach. Last but not least, a weak and not very powerful learner like PART achieved promising results with the new proposal based on a consistency measure and is able to compete with the best configuration of C4.5. To sum up, bearing in mind the new approach, for roc measure PART classifier with a consistency metric behaves slightly better than C4.5 and a correlation measure