Earthquake_UCLM

Based on aspects of building location and construction, your goal is to predict the level of damage to buildings caused by the 2015 Gorkha earthquake in Nepal.

---

ACTUAL SCORE: 0.7463 (647°)

• 0.7463 (updated 23/11/2025)
• 0.7461 (updated 23/11/2025)
• 0.7418 (updated 6/11/2025)

---

✅ Project Workflow - Pipeline (Expected to be carried out)

Exploratory Data Analysis (EDA)

• Perform in-depth EDA to understand feature distributions and relationships

• Visual tools to include:

  • Histograms
  • Boxplots
  • Pairplots
  • Correlation heatmaps

---

Outlier Detection

• Detect anomalous records that may deteriorate model performance

• Methods considered:

  • Isolation Forest

• Evaluate handling strategies (remove, cap, or model-based handling)

---

Baseline Model

• Implement first baseline using Random Forest

• Evaluate:

  • Cross-validation scores
  • Feature importance
  • Confusion matrix

---

Model Benchmarking

Test multiple gradient-boosting models:

• XGBoost
• LightGBM

Monitor and compare:

• Metrics micro-F1 main one but consider also: (accuracy, F1, etc.)
• Training time
• TRY TO AVOID Overfitting!!

---

Model Ensembling / Stacking

• Combine multiple models to boost prediction performance

• Initial plan:

  • Stack: XGBoost + LightGBM

• Explore blending strategies as well

---

Explainable AI (XAI)

• Assess model interpretability using:

  • SHAP - LIME
  • Feature importance
  • Partial dependence plots (optional)

---

Experiment Tracking

• Try MLflow to:

  • Track experiments
  • Log metrics
  • Register models
  • Compare performance

---

📂 Repository Structure (work in progress...)

├── Dataset/
├── Personal/
├── src/
│   ├── eda/
│   ├── models/
│   ├── utils/
├── mlruns/
├── README.md
└── requirements.txt

---

✅ Team Members

• Matteo Amagliani
• Ece Mina Örenler
• İrem Batıgün
• Kurbonmurodov Sardor
• Yazan Mousa

---

📌 Next Steps and notes

• If we have time, try more Feature engineering
• Test differents: Hyperparameter tuning
• Kaggle-style submission automation (idk if it's possibile on DataDriven)

---