Predicting Student Academic Performance Based on Psychological Test using Machine Learning

Abstract

It is essential to consider the psychological aspect of selecting new students to determine the success of prospective students. The psychological aspect is measured by a psychological test that shows the level of prospective students' abilities in social, emotional, personality, and potential to live at university. This paper proposes an approach to predicting student performance based on their psychological test scores using the Decision Tree and Random Forest algorithms. The dataset used in this study was taken from the student academic record at Institut Teknologi Del, which includes years of psychological test scores and the Grade Point Average (GPA) from studying at the Institute. More specifically, the dataset used includes the 2019, 2020, and 2021 class years. However, there are gaps in the dataset used, including missing values and psychological test attributes such as TIU, TIU Category, Work Achievement, Work Tempo, Accuracy, and Consistency, which are unavailable in other datasets. This is shown in the correlation heatmap, which shows the level of correlation for each attribute, which is still classified as a very weak correlation. Therefore, we came up with two approaches. The first approach is to use as many records as possible (Analysis on records), and the opposite of the second is to take advantage of more features (Analysis based on features). The two approaches are compared to determine which performs better for the classification model. Our results show that studies that emphasize the use of records produce slightly better performance than analyses that emphasize features. In more detail, the random forest algorithm produces the best performance compared to the decision tree algorithm in each Analysis, the RMSE value is 0.4552, and the MAE value is 0.3514. Moreover, none of the psychological test attributes strongly correlate to GPA and hence do not guarantee student performance.