Optimalisasi Prediksi Dalam Kelulusan Berbasis Deep Learning: Perbandingan Kinerja Multi-Layer Perceptron dan Deep Neural Network

Abstract

Predicting on-time graduation is one of the significant challenges in education, aiming to model the factors influencing academic success. This study aims to compare the performance of two Deep Learning algorithms, namely Deep Neural Networks (DNN) and Multi-Layer Perceptron (MLP), in predicting on-time graduation. The methodology used involves evaluating both algorithms with various performance metrics, including Recall, Accuracy, Precision, AUC, MCC, and Cohen Kappa. The results show that DNN performs better in terms of Recall (0.9766), indicating its ability to capture most of the students who graduate on time, although its AUC (0.8625) and Precision (0.8803) are lower compared to MLP. On the other hand, MLP excels in Accuracy (0.8812) and Precision (0.9037), providing more stable results for MCC and Cohen Kappa, demonstrating a better balance in predicting students who graduate on time and those who do not. Overall, while DNN is more sensitive in capturing students who graduate on time, MLP performs better in terms of balance between accuracy and minimizing prediction errors. This study suggests using MLP if the primary priority is accuracy and prediction stability, while DNN is more suitable when the main focus is capturing as many students as possible who graduate on time.