From Application to Graduation: A Fair Approach to Predict HEI Admissions and Academic Performance
Year: 2025
Type: dissertation
Abstract: The growing availability of data in higher education institutions (HEIs) has fostered the development of the Educational Data Mining (EDM) field. Among other advantages, EDM can enhance the admission process by supporting selection committees in making more accurate and impartial decisions. It also allows HEIs to monitor students’ academic performance, commonly depicted by their grade point average (GPA), and use this information to define personalized academic plans to help students according to their identified needs. However, the implementation of EDM solutions is not without risks, with one of the most pressing concerns being the potential presence of gender biases that inadvertently promote and perpetuate discriminatory practices within an academic context. In this study, we investigate whether EDM techniques can be used to predict student admissions and academic performance using exclusively application data. Simultaneously, we evaluate the presence of gender bias in these predictions and identify the variables that most contribute to each outcome, ensuring an interpretable and explainable solution. We utilized data from applicants to Nova IMS’s postgraduate and master’s programs between January of 2018 and November of 2024, testing a total of nine algorithms to predict HEI admissions (as a binary classification problem) and eight algorithms to predict the final GPA scores of the accepted students (as a regression problem). To interpret our results, evaluate the presence of gender disparities, and assess the influence of each independent variable on the predictions, we applied the Shapley Additive Explanations (SHAP) technique to the selected models. The results showed Adaptive Boosting (AdaBoost) to be the most effective algorithm for predicting admissions, reaching an F1 score very close to 0.9. On the other hand, Gradient Boosting (GBoost) was identified as the best algorithm to predict final GPA scores, with a root mean squared error (RMSE) of approximately 1.17 GPA units. For both tasks, the analysis of feature importance did not show evidence of perpetuating gender-based discrimination, opening the door to the implementation of this solution as a decision-support tool for admission committees in selecting candidates, and for academic services in designing student-tailored academic plans.
Author: Alves, Hugo Henrique Antunes
Institution: Universidade Nova de Lisboa
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112101717
