FAA Wildlife Strike Damage Prediction
Built a machine learning pipeline to predict whether FAA-reported wildlife strikes would cause aircraft damage using historical aviation incident data. Focused on imbalanced classification, feature engineering, and improving balanced accuracy across damage classes.
What it is
CS 363M team project: predict whether an FAA-reported wildlife strike damaged the aircraft. We had ~307k labeled incidents (1990–2015), ~34k unlabeled test rows, and a ~6.4% damage rate.
How it works
We cleaned messy mixed-type fields, kept sparse but safety-relevant variables such as speed and height, and treated missingness as signal. Domain features covered flight phase, wildlife size and count, and weather. We converted free-text remarks into structured features by searching for damage-related terms, explicit no-damage language, and inspection-only phrases. Each remark was compiled into binary flags and a damage keyword score, after which the raw text was dropped. High-cardinality categoricals (species, airport, operator) used smoothed out-of-fold target encoding so labels could not leak across folds. We compared a class-weighted logistic baseline, gradient boosting (XGBoost, LightGBM, CatBoost), a PyTorch MLP, then stacked with a logistic meta-learner. Evaluation used balanced accuracy and threshold tuning, not raw accuracy.
Outcomes
- Reached a private leaderboard score of 0.87, placing 16/50.
- Achieved a final balanced accuracy of 0.86 and placed 20/50.