Machine Learning to know if you R coming
summary
Models predict cancellations by combining survival analysis on booking lead time with feature engineering from channel, package type and inferred country (email/phone mining). Training uses H2O random forest and experiments with ranger; reported metrics are ~71% accuracy, AUC ~0.616 and recall ~0.32, with probability aggregation outperforming classification after two weeks. Production aggregates daily probabilities to decide releases and overbooking and achieves freeing ~16% of reservations ≈ 60% revenue when applied with risk thresholds and ethical safeguards.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:26Predicting cancellations
- 3:21Data sources and modeling
- 8:48Survival analysis results
- 10:58H2O AutoML and XGBoost
- 13:47Feature engineering and performance
- 15:34Risk-based reservation
- 17:47Ethical overbooking
- 22:00Q&A
related talks
24mPredicting Hotel Cancellations with Machine Learning
44mMachine learning @ booking.com
40mMachine Learning
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
30mDriving a Higher Level of Customer Support with Machine Learning
34mForecasting Airline Passengers Using Designer Machine Learning
22mPredict the future (or don't) with machine learning in applications
31mAppDev with Machine Learning and AutoML