Application of Machine Learning Models for Predicting School Dropout in Students from a Colombian Competency-based Education Institution

Abstract

Student dropout is a structural challenge in Colombian higher education, particularly in contexts with rigid curricular and pedagogical systems where the implementation of timely preventive strategies is complex. This study develops and validates a hybrid machine learning model, based on the CRISP-DM methodology, that integrates supervised algorithms (Random Forest, Ridge, XGBoost, KNN) and unsupervised approaches (K-Means, DECLA), supported by dimensionality reduction and segmentation techniques (PCA, MCA). Using sociodemographic variables, academic performance indicators, and a specifically designed monitoring instrument, the models achieved high accuracy in anticipating dropout risk and segmenting students into profiles of high, medium, and low probability of withdrawal. Tree-based algorithms, particularly Random Forest, demonstrated the best performance, identifying critical predictors such as number of complaints, grade reversals, socioeconomic status, gender, and marital status. The main contribution of this work lies in moving predictive analytics from an experimental exercise to an institutional support system in competency-based higher education, where academic rigidity often limits early interventions. By anticipating dropout through real-time empirical evidence, the model enables the design of differentiated action pathways personalized tutoring, socioeconomic support, and curricular flexibility that complement long-term educational reforms. In this way, its relevance in higher education is justified as an innovative and evidence-based resource to strengthen student retention