Abstract:

Academic dropout remains a persistent and extensively documented challenge in higher education, constituting a central research problem within Educational Data Mining and closely related domains such as Learning Analytics. Within this context, the prediction of student attrition is conceptualized as a knowledge discovery task, aimed at extracting actionable patterns from large-scale educational datasets to support evidence-based institutional decision-making. This study examines the application of supervised machine learning techniques within a data-driven analytical framework, focusing on the statistical estimation and early identification of student dropout risk. The analysis is conducted on a large and information-rich dataset comprising 56,492 instances and 64 predictive attributes, enabling the exploration of complex, non-linear relationships among academic, behavioral, and demographic variables. In line with established data mining pipelines, the preprocessing phase incorporates multiple stages of data preparation and transformation. These include statistical imputation methods for handling missing values, categorical variable encoding via one-hot encoding, feature scaling through z-score standardization, and feature engineering to enhance the dataset’s representational capacity. Such procedures are essential to ensure data quality and to optimize the performance of subsequent predictive models. Within this framework, three supervised classification approaches widely adopted in the literature are evaluated: K-Nearest Neighbors combined with a Bagging ensemble strategy, a feedforward Multilayer Perceptron with a simplified architecture, and the LightGBM model. These algorithms are selected due to their complementary inductive biases and their demonstrated effectiveness in extracting patterns from structured educational data. Model performance is assessed using balanced accuracy and recall, metrics particularly well-suited for imbalanced classification scenarios commonly observed in dropout prediction tasks. The experimental results demonstrate the robustness and effectiveness of data-driven approaches in modeling student dropout. The Multilayer Perceptron achieved a balanced accuracy of approximately 93%, while the KNN model with Bagging yielded competitive performance at around 87%. The LightGBM model outperformed the others, reaching approximately 94% balanced accuracy. These findings indicate that the evaluated models successfully capture latent structures in the data and achieve a satisfactory bias–variance trade-off within the adopted validation framework. Importantly, this study also highlights the role of model interpretability in data-driven educational research. To this end, Shapley values are employed to provide post hoc explanations for the predictions generated by the Multilayer Perceptron model. This approach enables the identification of the most influential features, as the student's workload, and allows for a fine-grained quantification of each variable’s contribution at the individual observation level. Such interpretability mechanisms are essential for translating predictive performance into actionable insights, thereby reinforcing the strategic value of data-centric approaches in understanding and mitigating academic dropout.

Keywords:

Processing techniques, dropout prediction, machine learning.