Enhancing Scholarship Allocation Fairness Using GRU and Synthetic Minority Oversampling Technique

Abstract

Fair scholarship distribution plays a crucial role in advancing educational equity; however, manual selection processes often suffer from subjectivity, inconsistency, and potential bias. To address these limitations, this study proposes a predictive framework that integrates the Gated Recurrent Unit (GRU) neural network with the Synthetic Minority Oversampling Technique (SMOTE) for classifying scholarship eligibility among high school students. The dataset consists of 367 student records obtained from SMAN 2 Makassar, encompassing academic and socio-economic attributes. Data preprocessing involved cleaning, categorical encoding, normalization, and stratified data splitting to ensure balanced representation. The baseline GRU model trained on imbalanced data achieved a validation accuracy of 91.38% and a test accuracy of 79.45%, revealing overfitting and bias toward majority classes. After the application of SMOTE to balance the minority class, the retrained model achieved improved performance with 96.55% validation accuracy and 83.56% test accuracy, accompanied by a significant reduction in test loss. These findings demonstrate that SMOTE effectively enhances the GRU model’s generalization capability and predictive fairness. Overall, the proposed GRU–SMOTE framework offers a robust, data-driven, and efficient approach for fair scholarship eligibility prediction. It can serve as a reliable decision-support system in educational data mining, promoting objectivity and transparency in scholarship selection processes.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint