Redes neuronales gráficas para predecir el rendimiento académico en Colombia

Abstract

The Colombian education system faces persistent challenges in equity and quality, particularly in rural and coastal public institutions (Suaza-Medina et al., 2024). In this context, the Saber 11 exams administered by ICFES constitute the main instrument for assessing academic performance at the end of secondary education. However, most existing predictive models rely on tabular approaches that treat students as independent observations, overlooking the interdependencies among actors, institutions, and territories. This study proposes a relational model based on Graph Neural Networks (GNN) to predict the overall score of the Saber 11 exams. The research employs the ICFES 2022-4 cohort (412,311 records and 35 variables), applying multiple imputation (MICE), normalization, dimensionality reduction, and the construction of an educational graph through cosine similarity and k-nearest neighbors (k- NN). Four GNN architecturesGCN, GAT, GIN, and GraphSAGEwere evaluated under a supervised regression scheme. Results indicate that GraphSAGE achieved the best performance (R2 = 0.987; MAE = 0.054; RMSE = 0.071; MedAE = 0.042), significantly outperforming traditional models. The variable importance analysis revealed that organizational and academic dimensions exert the strongest influence, followed by family, demographic, and institutional factors. These findings confirm that academic performance is a relational phenomenon shaped by interactions among students, institutions, and contexts. This study provides empirical evidence supporting the relevance of GNNs in education and lays the foundation for the development of early diagnostic and prediction systems that strengthen evidence-based and equitable educational decision-making.