Spacial Information Improves Modeling Accuracy in Predicting Student Performance

Abstract

The relationship between spatial heterogeneity and students' academic performance is a widely researched topic. This study extends the concept of spatial heterogeneity into two categories - allocatable (acquired after birth) and ascribed (determined before birth) - using seating and hometown information as variables in Spatial Durbin Models for statistical analysis, identifying seating as the salient feature. Subsequently, through 2,000 iterations of random state testing with undergraduate data from two classes (N=282), our research compares six representation learning methods, including an improved, custom version of attention-based Graphic Convolutional Network, affirming the superior efficacy of our enhanced approach in extracting seating features. By doing so, our study broadens the application scenarios of social comparison theory and group dynamics theory, while also providing a feasible best-practice reference for network-based instructional modeling.