Deep Learning for Scholarship Allocation: Hybrid CNN-FFNN Architecture

In this paper, we explore the application of different neural network architectures to a classification problem related to the allocation of graduate scholarships. Classical predictive models face limitations when it comes to simultaneously capturing the multidimensional complexity of data and managing class imbalance. We use classical models and artificial neural networks. We compare performance of classical models and artificial neural networks applied to predicting student eligibility for scholarships, in order to identify the most robust model best suited to the context. This comparison makes it possible to evaluate the suitability of different neural architectures for a hybrid dataset. To address these challenges, we propose an optimization approach by combining a CNN model, capable of efficiently exploiting sequential data, and a FFNN, adapted to static tabular data. The integration of the SMOTE technique also makes it possible to rebalance the classes, ensuring better robustness and generalizable performance. Experimental results show that this hybrid approach significantly improves the stability and accuracy of predictions compared to traditional models. This research contributes to increasing the accuracy and robustness some neural network model within the framework of an optimized decision support system.