Supplementary Materials for Included Study: Visibility Nowcasting in South Korea: A Machine Learning Approach to Class Imbalance and Distribution Shift

This repository provides the supplementary materials for the research article "Visibility Nowcasting in South Korea: A Machine Learning Approach to Class Imbalance and Distribution Shift." The included datasets and artifacts support full reproducibility of the experimental results, covering data augmentation, hyperparameter optimization, and distribution shift analysis.Contents1. `hyperparameter_optimization.pdf` - Detailed documentation of the hyperparameter search spaces and optimization setups for all predictive models (XGBoost, LightGBM, ResNet-like, FT-Transformer, DeepGBM) and the CTGAN data augmentation model.2. `oversampling_models_hyperparameters.csv` - Optimal hyperparameters identified for the SMOTENC and CTGAN oversampling models used to address class imbalance.3. `best_params_per_models` (directory) - CSV files containing the finalized optimal hyperparameters for each of the five predictive models across six regions (Seoul, Busan, Incheon, Daegu, Daejeon, Gwangju) and three cross-validation folds.4. `umap_plots` (directory) - High-resolution UMAP visualization plots (.png) comparing the data distributions of the original training sets with those of the augmented datasets (SMOTENC, CTGAN, Hybrid) for all regions and folds.5. `shap_plots` (directory) - Complete SHAP summary plots and feature importance values for all models, providing detailed interpretability analysis beyond the key features discussed in the main text.Usage NotesThese supplementary files are intended to be used in conjunction with the source code available at: (https://github.com/singbong/Visibility_Nowcasting)