Implementation of Feature Selection in Online Learning Performance Prediction Model

Online Learning has emerged vastly which resulted from Covid-19 pandemic. During the online learning period, several data were gathered regarding the performance of academic learning. The objective of this study is finding the most affected factor that resulted in the performance of the online learning by using feature selection method to determine the most important feature in predicting the result in a predictive model. The data was collected using a questionnaire in total of 361 respondents, which was utilized to develop machine learning models aimed at forecasting academic performance. In this study, several feature selection methods were implemented, including Pearson Correlation, Chi Square, Forward Selection, Backward Selection, Recursive Features, LASSO Regularization, and Random Forest Importance. The classifier models were built by using Random Forest and Bagging Decision Tree. Based on the feature selection methods that have been implemented, the top features are Performance (Without SMOTE) and Supporting Material (With SMOTE). The experimental findings showed that the best classifier was Bagging Decision Tree without SMOTE. The top three feature selection methods that achieved better F1-scores were Chi Square $(15\%)$, Random Forest Importance $(11\%)$, and Recursive Features $(7\%)$.