๐ง Daily Study Log [2025-08-13]
Focused on competition model refinement and documented a new urban safety project idea.
Achieved the best public LB score to date in the electricity forecasting competition.
Outlined a concept for an AI-powered urban flood prediction and mitigation system to minimize damage from sudden heavy rainfall.
The system would integrate real-time weather data, urban drainage capacity, and risk mapping to provide early warnings and optimize city responses.
๐ View idea
Conducted experiments with building-typeโspecific XGBoost modeling:
| No. | Description | Local SMAPE | Public LB SMAPE |
|---|---|---|---|
| 34 | Reduced overfitting by adjusting parameters: eta, subsample, colsample_bytree, min_child_weight, gamma, reg_alpha, reg_lambda. |
4.618441972505276 | 7.3385022051 |
| 35 | Split data by building type, optimized XGBoost hyperparameters via RandomizedSearchCV, applied 7-Fold OOF CV, then fixed once for stable performance evaluation. |
4.057732 | 7.3767868422 |
Best Score: ๐ 7.3385022051 (Experiment 34, Public LB)
๐ Urban Flood Shield: Concept for AI-driven flood prediction and prevention
๐ Refined XGBoost parameters to reduce overfitting โ best public LB score achieved
๐ Applied RandomizedSearchCV + OOF CV for stable building-typeโspecific model performance