Comparación de métodos de balanceo para la predicción de depresión con algoritmos de aprendizaje automático

Abstract

Abstract: Studies on depression make it possible to identify data records, data balancing techniques, models, and variables that are relevant for determining whether a person is experiencing a depressive state. Therefore, it is essential to carry out this type of research in order to analyze and determine which factors are associated with depressive symptoms. This thesis aims to compare data balancing methods for depression prediction using machine learning algorithms, following the CRISP-DM methodology. A review of related work is conducted to define the data balancing techniques, the best-performing models, and the metrics used for model evaluation. Data balancing techniques (oversampling, undersampling, and hybrid approaches) and supervised learning algorithms (Decision Trees, Random Forest, Gradient Boosting, XGBoost, LightGBM, AdaBoost, and CatBoost) are implemented, with hyperparameters tuned using GridSearchCV. The results are compared using evaluation techniques (confusion matrix, accuracy, precision, recall, F1-score, and AUC-ROC) to identify that the best-performing models were Random Forest, XGBoost, and LightGBM.