Clasificación del Desempeño Académico en la Prueba Saber 11 Mediante Variables Socioeconómicas: Estudio Comparativo Entre Redes Neuronales, Convolucionales y Sistemas Expertos

Abstract

The research work was developed as a comparative study of different artificial intelligence models ("MLP – CNN 1D - LightGBM") for the classification of academic performance based on the results of the Saber 11° State examination in Colombia, through socioeconomic, family, and institutional variables. It addresses the problem of the limitations in traditional statistical approaches when capturing the complex, non-linear, and high-dimensional structures of the variables that explain educational gaps in Colombia. To this end, multiclass predictive models were formulated to utilize a historical dataset of approximately 7.2 million anonymized records provided by the ICFES, corresponding to the periods from 2014-2 to 2024-2. The CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology was employed, implementing a hybrid processing pipeline and comparative analyses of the three classes of artificial intelligence models. The experimental results demonstrated that deep neural network architectures ("MLP") achieved the best predictive performance and highest statistical stability, outperforming the one-dimensional convolutional neural network ("1D CNN") and the ensemble algorithm (LightGBM) by capturing the non-linear dependencies of the dataset for each individual. In conclusion, socioeconomic variables possess a significant predictive capacity regarding academic performance classification, and deep learning models serve as a technological and methodological tool to validate the design of an early warning system and support evidence-based decision-making in the Colombian educational sector.