HRP Stage 01 — Booking Cancellation Risk

A trained LDA (Linear Discriminant Analysis) probability model for educational prediction of hotel booking cancellation risk, plus a separately saved Logistic Regression classifier. Trained with synthetic HRP booking records; not production-validated.

Task: Given 42 booking-time features, estimate the probability of later cancellation. This is decision support only: do not automatically decline or cancel bookings based on this score.

Model selection and verified source scores

From the original results/final_test_metrics.json:

Model	Selection basis	Test F1	Test ROC-AUC
LDA probability model, tuned threshold 0.265	Validation ROC-AUC	0.6068	0.7234
Logistic Regression classifier	Validation F1	0.5866	0.7228

The LDA pipeline applies its own fitted preprocessing internally, including imputation, scaling and one-hot encoding.

Synthetic training split: chronological 70/15/15, 15,000 rows total. The public companion Dataset has 42 feature columns, a non-input created_at, and target_cancelled.

Files

models/risk_probability_model.joblib: primary fitted LDA scikit-learn Pipeline.
models/best_classifier.joblib: secondary fitted Logistic Regression Pipeline.
models/preprocessor.joblib: separately stored preprocessing artifact from the original package; not applied a second time in src/predict.py.
models/model_metadata.json: 42 features and risk threshold.
src/predict.py: input schema handling and prediction function.
integration/request_example.json: complete synthetic 42-feature example request.
results/final_test_metrics.json: original metrics.
requirements_inference.txt: minimal inference dependencies.
example_inference.py: runnable example.
examples/risk_examples.json: three synthetic holdout examples spanning LOW / MEDIUM / HIGH risk bands.

Use in Google Colab / local Python

Load only trusted .joblib files: joblib deserialization can execute Python code.

pip install -r requirements_inference.txt
python example_inference.py
# Or:
python -m src.predict integration/request_example.json

No retraining is required for inference. Input must contain the 42 model features exactly as named in models/model_metadata.json; extra keys are ignored, but missing required keys raise an error. created_at and target_cancelled are not model input features. The output includes cancellation probability, binary risk flag at 0.265, and LOW (<0.35) / MEDIUM (<0.65) / HIGH risk levels. A thresholded risk flag is not a guarantee of future cancellation.

Limitations and ethics

This is an academic prototype trained on synthetic HRP-shaped bookings. It does not establish accuracy, fairness, or calibration on real-world bookings.
No real HRP customer records are included in the released companion dataset.
Do not use for automatic cancellation, rejection, customer penalties, or other adverse decisions.
Re-train and evaluate on consented, verified booking-time snapshots and realized outcomes before production.
Model binary artifacts should be loaded only in trusted environments.

Group collaboration

Student contributions should be submitted from each student's individual Hugging Face account using PRs/discussions/commit history; do not claim contributions that are not visible in repository history.

Downloads last month: -; Downloads are not tracked for this model. How to track