Early prediction Students' Graduation Rank Using LAGT: Enhancing Accuracy with GCN and Transformer on Small Datasets

Abstract

Recent efforts to predict students' graduation ranks using machine learning and deep learning methods have faced challenges, particularly with small sample sizes which limit accuracy. This paper introduces the LAGT (Learning Analysis by Graph Convolutional Network and Transformer) method, a novel approach for early predicting of students' graduation ranks. LAGT integrates a Graph Convolutional Network (GCN) to enhance the training set with labeled samples and utilizes a Transformer to forecast graduation ranks. This method harnesses the semi-supervised learning capabilities of GCN to automatically label data, addressing the constraints of small sample sizes in training sets. Additionally, the Transformer leverages its proficiency in handling long sequences and capturing contextual information, thereby demonstrating superior effectiveness in models trained on larger datasets. We evaluated this method on three datasets from some universities (HNMU1, HNMU2, VNU) and achieved a maximum accuracy of 92.73%. Results indicate that the integrated LAGT method outperforms comparable approaches across multiple metrics including accuracy, prediction precision, and model sensitivity, achieving up to a 35.73% improvement. Notably, on the same HNMU1 dataset, the accuracy increased from 85% (reported by Son et al.  [1] to 90.91% with this model. Experimental comparisons underscore the superior performance of LAGT over alternative methodologies in similar scenarios.