PREDICTING COURSE GRADES OF STUDENTS’ ACADEMIC PERFORMANCE USING THE LIGHTGBM REGRESSOR

Abstract

Abstract. In the modern world, using all available opportunities and technologies, special attention should be paid to the development of the education system of students, since education serves as the basis for the development of the future generation. Nowadays, thanks to the use of available Artificial Intelligence methods, it is possible to predict various events, anomalies or other important things. With the help of machine learning, it is possible to predict at an early stage of a student's education whether he will finish the course successfully or not. In this study, it is proposed to predict the final score which student will receive at the end of the course using a number of predictors as an assessment for the first quiz and 3 types of tasks using the LightGBM regressor, which is a high-performance algorithm with gradient boosting. The results of using the LGBM regressor using GridSearchCV allowed  to determine the best settings of hyperparameters from three selected tree-like boosting methods: 'dart', 'gbdt', 'goss'.  The GOSS method was determined to be the best of the three methods listed with a standard error for the test set of (-5.76) and an estimate of the accuracy of the model forecast in 81%, which is 24% more than the accuracy of the Linear Regression forecast of – (57%).
    Keywords: Machine learning, grades prediction, outliers’ identification, LGBM Regressor, Linear Regression.