Implementation of Feature Selection Based on Particle Swarm Optimization and Genetic Algorithm on Support Vector Regression Algorithm to Predict Student Performance

Abstract

Identifying student learning outcomes data in a subject to predict student performance is interesting research. This study was conducted to identify the factors that affect student performance by applying feature selection techniques based on Particle Swarm Optimization (FSPSO) and feature selection techniques based on Genetic Algorithm (FSGA) on the Machine Learning algorithm, namely Support Vector Regression (SVR). The data used are students' personal data and data on the academic value of Mathematics in Portugal. Experiments were carried out by applying the PSO feature selection (FSPSO) and GA (FSGA) feature selection techniques to obtain the selected features. Then the prediction modeling of student performance was carried out using SVR. The experiment was carried out using the pycharm 2021.1 application with the python programming language. This study uses a 10-fold cross-validation method. Evaluation of student performance prediction modeling is done by comparing the error values of regression performance metrics and processing time with feature selection and without feature selection. The experimental results show that FSPSO-SVR modeling compared to SVR modeling, the error reduction of the regression performance metric on RMSE is 34.36%, while for MAE it is 45.09% in 500 iterations. In FSGA-SVR modeling compared to SVR modeling, the error reduction of the regression performance metric in RMSE is 1.59% %, while the MAE is 3.04% at iteration 50. The improvement of the time process value in the FSPSO-SVR modeling compared to the SVR modeling, resulted in a decrease in process time of 44.68%. while the FSGA-SVR modeling compared to SVR modeling, resulted in a decrease in process time of 33.33%. From the comparison of experimental results, it can be concluded that the FSPSO-SVR modeling is the best predictive modeling of student performance with the lowest error value and the fastest time process in 500 iterations, so it can be used as a reference in making decisions to improve student performance in the future.