Evaluation of prediction models for slope stability on road cut slopes using XGBoost, LightGBM, and CatBoost

This study aims to develop a prediction model for slope stability using XGBoost, LightGBM, and CatBoost, based on slope investigation data. To thoroughly evaluate the variables affecting slope stability, feature importance is analyzed. The analysis is divided into three cases: Case-1, which includes numerical variables; Case-2, consisting of categorical variables; and Case-3, incorporating both numerical and categorical variables. The average prediction performance for the test models is approximately 67-68% in Case-1, 77-78% in Case-2, and 77-79% in Case-3. All models using XGBoost demonstrate significant differences depending on the output values and a trade-off between recall and precision. LightGBM exhibits the highest prediction performance in Case-1, while also showing a trade-off between recall and precision, as well as overfitting in Case-2 and -3. Although CatBoost tends to overfit the training model, it achieves the highest prediction performance, approximately 90%, in the training model of Case-3. All models across each analysis case and machine learning library exhibit excellent prediction performance and area variable for use. However, when precise results are needed, LightGBM is recommended for its balanced performance, and CatBoost is suitable when a highly accurate model is required.