j2damax/hotel-cancel-model

Hotel Booking Cancellation Predictor

Predicts probability that a hotel booking will be cancelled (Sri Lankan hospitality context). The champion model is XGBoost; threshold based decisions currently use 0.35000000000000003 (see champion_meta.json).

Last updated: 2025-10-05 16:43 UTC

Key Metrics (Holdout)

Metric	Value
F1	0.8046506137865911
ROC-AUC	0.9384035807110922
Precision	0.841708852944808
Recall	0.7707179197286602
Accuracy	0.8613786749308987

Top Features (SHAP importance)

• deposit_type
• country__te
• market_segment
• total_of_special_requests
• lead_time
• required_car_parking_spaces
• assigned_room_type
• customer_type_target_encoded
• reserved_room_type
• previous_cancellations

Quickstart

from huggingface_hub import snapshot_download
import joblib, json, pandas as pd

local_dir = snapshot_download(repo_id="j2damax/hotel-cancel-model")
model = joblib.load(f"{local_dir}/champion_model.pkl")
preprocessor = joblib.load(f"{local_dir}/preprocessor.pkl")
meta = json.load(open(f"{local_dir}/champion_meta.json"))

sample = pd.DataFrame([{
    'lead_time': 45, 'arrival_month': 7, 'adults': 2, 'children': 0, 'adr': 110.0
}])

X = preprocessor.transform(sample)
proba = float(model.predict_proba(X)[:,1][0])
print('Cancellation probability:', round(proba, 4))

Files

• champion_model.pkl – serialized champion estimator
• preprocessor.pkl – unified preprocessing / feature pipeline
• champion_meta.json – metrics & threshold
• Optional SHAP / feature importance JSON artifacts

Notes

Model trained with stratified 5-fold CV; primary selection metric: F1; tie-breaker: ROC-AUC. Class imbalance handled via class weights.

Citation

Academic coursework (NIB 7072) — Sri Lankan tourism cancellation risk analysis.