Airline passenger satisfaction
The score is only useful if the missing delays and the repeated survey answers are handled before a model is allowed to see them.
- XGBoost
- SHAP
- PCA
- imputation
- 0.987ROC-AUC, held-out flights
- 0.935F1 on that same split
- 0.93R² filling arrival delay
Repair the table first
Mean-filling a delay pretends every missing flight was ordinary. Predicting arrival from the departure delay that did happen keeps the delay in the row.
- 103,904 records. Missingness separated into MCAR, MAR, and MNAR before anything was filled.
- Arrival-delay imputation reached an R² of 0.93.
- Eighteen features started above a VIF of 5. Fourteen survey ratings went through PCA. After that, none of the remaining columns did.
A fair comparison, then an explanation
Once the repeated survey items are combined, a simple model and a boosted one can be compared on the same columns. The winner still needs a reason a person can read.
- Logistic regression, a random forest, and XGBoost, one feature set. XGBoost is the reported model.
- Held-out ROC-AUC 0.987, F1 0.935.
- Feature choice also compared mutual information, recursive feature elimination, and L1 logistic regression.
What the score is actually saying
Cabin matters, and the reason for flying matters with it. The same seat does not mean the same passenger.
- Business class is satisfied about 69% of the time, Eco 19%, Eco Plus 25%.
- Inside business class, a business trip is about 72% satisfied and a personal trip about 12%.
- SHAP puts online boarding and type of travel at the front of the explanation. Almost all disloyal customers in this table are business travelers, so retention here is a corporate-account question.