Implementasi Algoritma LIGHTGBM dan Xgboost Menggunakan Metode Semi-Supervised Learning pada Data Klasifikasi Kesehatan Mental

Abstract

Mental health is a critical issue, particularly among university students who are susceptible to both internal and external pressures. Early classification of the need for mental health treatment helps determine whether professional intervention is required. This study aims to develop a prediction model for mental health treatment needs using a semi-supervised learning approach with a pseudo-labeling technique. Pseudo-labeling is employed to integrate local population characteristics specifically from FMIPA University of Jember into the prediction model, thereby enhancing the relevance of the respondent data. A secondary objective is to compare the performance of two gradient boosting algorithms LightGBM and XGBoost. The study utilizes a public mental health dataset comprising 261.328 records. Preprocessing steps include removing irrelevant columns, handling missing values, and applying binary, ternary, and ordinal encoding, resulting in a dataset with 12 numerical input features. The data is split with (70:30) into training 70%, validation 15%, and testing sets 15%. A separate set of 100 questionnaire responses is used for the pseudo-labeling process, wherein predicted labels derived from the initial model’s patterns are integrated, and the model is retrained. The final model is the tested against the remaining 28 untouched questionnaire responses to evaluate its implementation. The results indicate that the final model achieves an accuracy of 70,70% and an AUC of 0,7580 with the LightGBM algorithm, while XGBoost achieves an accuracy of 70,65% and an AUC of 0,7545. Analyses of precision, recall, and F1-Score show nearly identical values for both algorithms. While the comparison reveals comparable predictive performance across all evaluation metrics, LightGBM is concluded to be superior to XGBoost due to its faster training time 2,2 seconds compared to XGBoost 6,6 seconds.