🚦 Road Accident Risk — Residual-Boosted Risk Model

Playground Series S5E10 (Kaggle)

🎯 What is this?

Predict accident_risk for each road segment as a calibrated score in [0,1].

Goal:

• Stable CV, not leaderboard luck
• Interpretable signals (why is this road risky?)
• Zero leakage

🧠 Modeling Pipeline (3 stages)

1. LightGBM (main learner)

• Train LightGBM directly on accident_risk
• Bag multiple random seeds → smoother OOF preds
• Output: oof_lgb, pred_lgb

2. XGBoost residual (prior-corrected)

• Build an interpretable safety prior risk_prior ∈ [0,1]

  • high curvature
  • high speed limit
  • night lighting
  • bad weather

• Train XGBoost on the residual:
  residual_target = accident_risk - risk_prior

• At inference:
  pred = risk_prior + predicted_residual

• Output: oof_xgb, pred_xgb

Why? Stage 2 is only learning what the simple prior missed.

3. NNLS blend (non-negative)

• Fit Non-Negative Least Squares (NNLS) on [oof_lgb, oof_xgb]
• Get blend weights ≥ 0 (no negative canceling)
• Apply same weights to test preds
• Clip final predictions to [0,1]
• Output: final_test → submission.csv

Result:

• Lower OOF RMSE
• More consistent folds
• Predictions always in a valid range

🔬 Features & CV

Feature engineering

• curv_speed = curvature × speed_limit
• acc_per_lane = num_reported_accidents / num_lanes
• critical_zone = high curvature & high speed
• risk_prior = human-readable baseline danger score

Cross-validation

• Stratified K-Fold on binned target quantiles
• Keeps each fold balanced (safe vs dangerous segments)
• All metrics are out-of-fold (OOF)

📂 Output

The notebook will:

1. Train Stage 1 → Stage 2 → Stage 3
2. Blend predictions
3. Write submission.csv under artifacts/ (Kaggle: /kaggle/working/artifacts/)

No external data. No test target leakage.

📁 Repo layout

.
├── road-accident-risk-ps5e10.ipynb
├── data/
│   └── raw/                     # (optional) place train/test CSVs here for local runs
├── artifacts/                   # saved outputs (e.g., submission.csv)
├── repo_utils/
│   └── pathing.py               # local data/raw + Kaggle /kaggle/input fallback
├── CASE_STUDY.md
├── requirements.txt
└── .gitignore

📦 Data loading (local + Kaggle)

The notebook resolves files in this order:

1. DATA_PATH env var (full file path)
2. Local data/raw/<filename>
3. Kaggle /kaggle/input/<dataset>/<filename>

For local runs, place these files under data/raw/:

• train.csv
• test.csv
• sample_submission.csv

🚀 Run locally

python -m venv .venv
# Windows: .venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt

Open road-accident-risk-ps5e10.ipynb and run top-to-bottom.

📤 Outputs

• artifacts/submission.csv (Kaggle: /kaggle/working/artifacts/submission.csv)