Airline Delay Root-Cause Modeling
Built an end-to-end machine learning pipeline to model and explain flight delays across the U.S. using multi-year airline operations (BTS), NOAA weather data, and airport activity metrics (~800k records).
What I built: • Integrated heterogeneous datasets (airline ops, weather, airport congestion) • Engineered ~44 features capturing time-of-day, route, and environmental effects • Trained and evaluated Logistic Regression, Random Forest, and Gradient Boosting • Performed threshold tuning to analyze precision–recall tradeoffs under class imbalance (~80/20)
Key results: • Random Forest provided the best balance (F1 ≈ 0.40, recall ≈ 0.58) • Logistic Regression achieved highest recall (~0.62) • Gradient Boosting had strongest ranking (ROC-AUC ≈ 0.70) but low recall at default thresholds
Key insight: Flight delays are driven primarily by system-level factors, not isolated events. Time-of-day (network congestion) and weather (precipitation, wind) consistently dominated across models.
Takeaway: This project reinforced that effective ML is not just about accuracy—it’s about understanding tradeoffs and extracting actionable insight from complex systems.