Enhancing Student Performance Prediction with Graph Representation Learning Based on Gower Distance

In the field of educational data mining, predicting student performance is both crucial and complex. This study introduces a novel method combining Gower distance and Graph Neural Networks to enhance the accuracy of predicting student academic outcomes. Initially, we calculate the similarity between student samples using Gower distance, a metric capable of effectively managing the mixed data types (numerical and categorical) present in student performance data. Based on these calculated similarities, we construct a graph dataset of students, where nodes represent students and edge weights denote the similarities between them. Within this graph dataset, we deploy an improved GowerSAGE model to predict student performance. Our enhancement focuses on sampling neighbors during the neighbor sampling process, specifically choosing those in close proximity to the target student in terms of Gower distance. This strategy ensures the model focuses on samples similar in performance to the target student, thereby more accurately capturing and utilizing the interrelations and similarities among students. Experiments conducted on three public datasets demonstrate the superiority of our model in predicting student performance.