Analysis of Influencing Factors in Predicting Student Academic Achievement in Rural and Urban Areas of Eastern Indonesia using a Graph Neural Network and Shapley Additive Explanations

Abstract

The quality of education in Indonesia remains uneven, particularly between rural and urban areas.
This study aims to model National Assessment (AN) data into a graph structure and analyze
the differences in factors influencing the prediction of student achievement scores using Graph
Neural Network (GNN) and SHapley Additive exPlanations (SHAP). The data used involves
2,162 students (1,329 rural and 833 urban). The results show that the data was successfully
represented as a heterogeneous graph with 3,318 nodes and 4,403 relations. The GNN model
achieved an optimal test accuracy of 70.37%. SHAP analysis reveals that the model’s prediction
decisions are dominated by school institutional variables, followed by socioeconomic variables.
Class density (Number of Students per Class) is the most crucial determinant in both areas (mean
|SHAP| rural = 0.0671; urban = 0.0864). However, there is a difference in the priority of the
subsequent dominant factors; in rural areas, economic barriers (Number of PIP Students, mean
|SHAP| = 0.0385) become the second main driver. Conversely, in urban areas, institutional scale (Number of Students, mean |SHAP| = 0.0420) is more dominant. Opposing impact direction
patterns were also found in certain variables such as the number of students. In conclusion,
regional ecosystems create structurally different dynamics of influence among variables, meaning
educational policy interventions must be tailored to specific regional challenges.