A Feature Selection Method Based on Rough Set Attribute Reduction and Classical Filter-Based Feature Selection for Categorical Data Classification

The main objective of feature selection in machine learning classification is to reduce the size of features by removing irrelevant and noisy features, with the goal of improving the accuracy and the efficiency of the classification model. Like continuous and mixed data classification, feature selection has been applied to better categorical data classification. On large datasets with tens of features, however, existing feature selection methods perform worse in terms of accuracy metrics than baseline categorical data classification models that involve full features. This paper presents a feature selection method that integrates Rough Set Attribute Reduction and Classical Filter-based feature selection method to improve the performance of categorical data classification. Two large categorical datasets from UCI repository are used to evaluate the method. Support Vector Machine, Random Forest and Multilayer Perceptron algorithms are used as machine learning classifiers. The results show that the proposed method outperforms existing feature selection models in terms of Accuracy, Precision, Recall, and F-measure for individual classes and their average weighted scores in both case studies. Benchmarking with baseline classification models, the best overall performance by the proposed method is obtained with Random Forest.