Diseño y desarrollo de modelo predictivo integral, explicable y multinivel para la detección temprana y el seguimiento del riesgo de deserción estudiantil

Abstract

This document presents the comprehensive development of a student dropout predictive model for the Universidad Distrital Francisco José de Caldas, conducted as an internship at the Advisory Office of Technology and Information (OATI). The project evolves the existing institutional model—built under the CADEP Acacia framework—by proposing a two-phase architecture: a Logistic Regression that estimates initial risk on enrollment day, and an XGBoost model using SMOTE for class imbalance correction that refines this estimation at the end of the semester. Evaluated on a historical dataset of 90,485 enrollment records, the new system achieves a ROC-AUC of 0.865 and a recall of 77.8% for actual dropouts, significantly outperforming the previous approach replicated on the same data (0.805 and 47.0%, respectively). The central contribution of this work lies in its explainability layer: using SHAP values, each prediction provides the student's risk level, the top three contributing factors to that risk, and their specific values. This enables the Institutional Welfare department to design targeted, tailored interventions rather than applying generic protocols. The system was rigorously validated through 58 automated tests and 13 formal functional tests, and includes a trend analysis of the historical data. Furthermore, it ensures compliance with Data Protection Law 1581 of 2012 by anonymizing all student identifiers. Finally, at OATI's request, the model was deployed as a service via a Dockerized REST API, featuring Swagger documentation, audited security, and zero-failure load testing, leaving it ready for migration to the institutional production repository.