Predicting Students’ Academic Performance: A Comparison between Traditional MLR and Machine Learning Methods with PISA 2015

Abstract

Predicting students’ academic performance has long been an important area of research in education. Most existing literature have made use of traditional statistical methods that run into the problems of overfitted models, inability to effectively handle large numbers of participants and predictors, and inability to pick out non-linearities that may be present. Regression-based ML methods that can produce highly interpretable yet accurate models for new predictions, are able to provide some solutions to the aforementioned problems. The present study is the first study that develops and compares between traditional MLR methods and regression-based ML methods (i.e. ridge regression, LASSO regression, elastic net, and regression trees) to predict students’ science performance in the PISA 2015. A total of 198,712 students from 60 countries, and 66 student- and school-related predictors were used to develop the predictive models. Predictive accuracy of the various models built were not that different, however, there were significant differences in the predictors identified as most important by the different methods. Although regression-based ML techniques did not outperform traditional MLR, significant advantages for using ML methods were noted and discussed. Moving forward, we strongly believe that there is merit for using such regression-based ML methods in educational research. Educational research can benefit from adopting ML practices and methods to produce models that can not only be used for explaining factors that influence academic performance prediction, but also for making more accurate predictions on unseen data.

Predicting VO2max in Children and Adolescents Aged between 6 and 17 Using Physiological Characteristics and Participation in Sport Activities: A Cross-Sectional Study Comparing Different Regression Models Stratified by Gender

“…After applying different regression models of different complexity, and by using performance indicators such as RMSE, MAE, and R 2 , we concluded that the classical MLR model performs better than more complex models, such as QR, RR, and RT, and competed with SVR-RBF, which outperforms it in all indicators in the boys’ model and in R 2 in the girls’ model and similar results were found in the literature [ 58 ]. These findings are consistent with other findings in the literature, as we have large-scale data with only a few outliers, with normality of residual distributions, and with acceptable levels of multicolinearity [ 59 , 60 ].…”

Using machine learning to predict UK and Japanese secondary students' life satisfaction in PISA 2018

“…Although ML models have shown great potential in predicting students' cognitive and affective outcomes, most of the existing studies focused on students' academic achievement and career choice, with little attention to their subjective well‐being (Dong & Hu, 2019; Mandalapu & Gong, 2019; Puah, 2021; Yeung & Yeung, 2019). Although several studies paid attention to subjective well‐being, few of them investigated the effects of student‐related variables, such as the teacher support and learning goals (Kaiser et al., 2021; Morrone et al., 2019; You, 2021).…”