Performance evaluation of chi-square and relief-F feature selection for facial expression recognition

Abstract

Pattern recognition is a crucial part of machine learning that has recently piqued scientists' interest. The feature selection method utilized has an impact on the dataset's correctness and learning and training duration. Learning speed, comprehension and execution ease, and properly chosen features influence all high-quality outcomes. The two feature selection methods, relief-F and chi-square, are compared in this research. Each technique assesses and ranks attributes based on distinct criteria. Six of the most important features with the highest ranking have been chosen. The six features are utilized to compare the performance accuracy ratios of the four classifiers: k-nearest neighbor (KNN), naive Bayes (NB), multilayer perceptron (MLP), and random forests (RF) in terms of expression recognition. The final goal of the proposed strategy is to employ the least number of features from both feature selection methods to distinguish the four classifiers' accuracy performance. The proposed approach was trained and tested using the CK+ facial expression recognition dataset. According to the findings of the experiment, RF is the best accurate classifier on chi-square feature selection, with an accuracy of 94.23 %. According to a dataset utilized in this study, the relief-F feature selection approach had the best classifier, KNN, with an accuracy of 94.93 %