Analysis of Student Academic Performance Using a Hybrid Autoencoder–Deep Neural Network (AE-DNN) Model and a Multilayer Perceptron (MLP)

Abstract

Student academic achievement serves as a key indicator for evaluating the success of the school learning process. Advances in artificial intelligence, particularly Deep Learning, enable more accurate and efficient analysis of educational data. This study aims to analyze and predict student academic achievement based on report card grades using Autoencoder-Deep Neural Network (AE-DNN) and Multilayer Perceptron (MLP) models. The dataset comprises grades from 10 subjects—Mathematics, Natural Sciences, Social Sciences, Indonesian, English, Civics, Physical Education, Islamic Religious Education, Arts, and ICT—for 320 students. The data underwent preprocessing steps, including cleaning, normalization, and a split into 80% training data and 20% testing data. The Autoencoder method was employed for feature extraction and dimensionality reduction to generate an optimal data representation prior to processing by the Deep Neural Network. Subsequently, the performance of the AE-DNN model was compared with that of the MLP model in classifying student academic achievement levels. Model evaluation was conducted using accuracy, precision, recall, and F1-score metrics. The results indicate that, in terms of overall accuracy, both architectures achieved the same score of 98.44%, correctly classifying 63 out of 64 test samples. However, category-specific F1-score analysis revealed complementary characteristics: the AE-DNN model proved superior and more precise in predicting the "High" category (F1=1.0000) compared to the MLP (F1=0.9630), whereas the MLP model performed perfectly in the "Low" category (F1=1.0000). This research is expected to assist schools in the early identification of student achievement levels, thereby serving as a basis for decision-making to improve the quality of instruction and student academic support.