STUDENT ACADEMIC PERFORMANCE ANALYSIS USING SUPPORT VECTOR REGRESSION AND MULTILAYER PERCEPTRON: AN EDUCATIONAL DATA MINING APPROACH

Abstract

For its ability to support data-driven educational decision-making, student academic performance prediction is an important effort in Educational Data Mining. The objective of this study is to evaluate the relative performance of SVR and MLP with each other on predicting student academic performance using an organized dataset that we collected from Kaggle containing 10,000 rows. The dataset contains learning-related features, including study hours, previous academic achievement scores due to exams taken previously (hereafter pre-study score), extra-curricular activities, sleeping time and intensity of practicing. Extensive data preprocessing was performed and binary transformation of categorical variables and normalisation of numerical attributes were employed to assure compatibility with the model. The models performance was tested with the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and coefficient of determination (R²). Learning curve analysis and actual versus predicted plots were used to investigate model generalisation and learning characteristics. The findings show that SVR and MLP can predict student academic performance with high level of accuracy. Nevertheless, SVR exhibits better stability for generalization, with smaller gaps between the training and validation errors and lower MAE/RMSE values. On the other hand, MLP has more flexible capability to capture complex learning patterns but with slightly larger variance in prediction errors. This research is relevant in the field of Educational Data Mining, through the comprehensive analysis of kernel and neural network-based regression models for academic performance prediction.