Experimental analysis of filtering-based feature selection techniques for fetal health classification

Abstract

Machine learning techniques enable computers to acquire intelligence through learning. Trained machines can carry out various tasks, such as prediction, classification, clustering, and recommendation, within a wide variety of applications. Classification is a supervised learning technique that can be improved using feature selection techniques such as filtering, wrapping, and embedding. This paper explores the impact of filtering-based feature selection techniques on classification methods, and focuses on an analysis of correlationbased filtering techniques based on Pearson, Spearman, and Kendall rank correlation. Similarly, we explore the impacts of using statistical filtering techniques such as mutual information, chi-squared score, the ANOVA univariate test, and the univariate ROC-AUC. These filtering techniques are evaluated by implementing them with the k-nearest neighbor, support vector machine, decision tree, and Gaussian na?ve Bayes classification methods. Our experiments were carried out using a fetal heart rate dataset, and the performance of each combination of methods was measured based on precision, recall, F1-score, and accuracy. An analysis of the experimental results showed that the performance metrics for the Gaussian na?ve Bayes and k-nearest neighbor methods were improved by 3% through the use of the statistical feature selection technique, and a 4% improvement was observed for the decision tree and support vector machine methods using a correlation-based filtering technique. Of the statistical feature selection techniques, ANOVA and ROC-AUC were the best as they improved the accuracy by 92%; compared to the other correlation techniques, the Spearman correlation coefficient gave the best results, as it also improved the accuracy by 92%.