Prediction of Academic Grades in Higher Education Using Deconvolutional Density Networks

Na Zhang, Shenghui Zhu, Shuren Li, Lin Wang, Bo Yang, Yuhu Cheng

Abstract: Ensuring successful graduation and reducing dropout rates are fundamental objectives of higher education. Many studies in the field of educational data mining (EDM) attempt to model the relationships between various student-related data and academic grades, aiming to identify students in need of assistance early and intervene promptly to prevent academic failure. However, privacy and cost issues limit the sufficient and effective collection of student data, making traditional single-value student grade predictions insufficient to capture the various uncertainties present in incomplete data scenarios. Therefore, researchers have turned to using Mixture Density Networks (MDNs) to predict student grades and generate an explicit Probability Density Function (PDF), providing more decision support information for educational administrators. However, MDNs often lack robustness and are sensitive to parameters in probability density estimation algorithms, which is not conducive to effective educational decision-making support. To address these issues, we adopt Deconvolutional Density Networks (DDNs) to model the explicit PDF of student grades. DDNs discretize grade ranges into multiple uniform bins and use piecewise constant functions to represent the probabilities within each bin, mitigating the challenges encountered by MDNs in predicting student grades. We conducted comparative experiments between DDN and MDN on three courses from two colleges within our institution. The experimental results demonstrate that DDN outperforms MDN and exhibits superior robustness.