Academic Performance Prediction Based on Convolutional Neural Networks and IRT Parameters as RGB Images

Abstract

In today’s competitive educational environment, institutions face the crucial challenge of effectively assessing student performance, a problem of utmost importance to ensure quality education and develop strategies that improve academic performance and anticipate future demands. The literature explores various approaches to predict student performance using Item Response Theory (IRT) parameters and machine learning techniques. However, there needs to be more in computer vision to capture the behaviour of question assertiveness in image form. This work proposes transforming the IRT parameters into RGB matrices to generate images, which are used to train a convolutional neural network model. The results demonstrate the effectiveness of this method, showing that the images corresponding to the highest scores have a lighter tone, reflecting a more significant number of correct answers and, consequently, greater pixel intensity. Furthermore, the model successfully learned the students’ scoring patterns, generating a Spearman Correlation for RGB Images of 0.86 for 20,000 images, showcasing its strong generalization capabilities.