🏨 Hotel Booking Cancellation — Customer Segmentation

Statistical Learning — Master 2 TIDE, Université Paris 1 Panthéon-Sorbonne Academic year 2025–2026 | Professor: Alain Celisse

---

📌 Project Overview

Analysis of hotel booking cancellations on the INN Hotels Group dataset (36,275 bookings, 19 variables). The project covers the full ML pipeline from EDA to business impact analysis, combining supervised and unsupervised approaches.

#	Research Question	Method	Author
Q1	Predicting cancellations — EDA, business impact	Logistic Regression · LASSO/Ridge · Random Forest	M. Hadmen
Q2	Discriminant analysis & boosting	LDA · QDA · KNN · XGBoost	B. Kessi
Q3	Unsupervised customer segmentation	GMM · PCA · BIC/AIC	A. Mattei

---

🔬 My Contribution — Q3: Customer Segmentation via GMM

Objective

Identify natural booking profiles in an unsupervised setting — without using the cancellation label at any point — to uncover hidden customer segments and their cancellation behaviour.

Methodology

• Preprocessing: one-hot encoding of categorical features, standardisation → 27-dimensional feature space
• Dimensionality reduction: PCA (10 components, 58.7% variance explained)
• Model selection: BIC/AIC minimisation over K ∈ [2, 8] → K = 4 (EM convergence in 24 iterations)
• Assignment: MAP rule on posterior probabilities

Results — 4 Customer Profiles

Cluster	Profile	Size	Cancellation rate
C0	Standard	31.5%	33%
C1	Upscale	34.7%	36% (highest)
C2	Early Planners	25.7%	33%
C3	Loyal Guests	8.2%	17% (lowest)

96.9% of observations have a cluster membership probability > 0.9, confirming well-separated clusters.

Visualisations

• BIC / AIC curve for K selection
• 2D PCA projection of GMM clusters vs actual cancellation status
• Cluster profiling: lead time, price, special requests, loyalty rate
• Membership probability distribution (MAP rule)

---

🏆 Key Results (Full Project)

Metric	Value
Best model	XGBoost
AUC-ROC	0.956
AUC-PR	0.935
Optimal threshold	0.35 → Recall = 91.8%
Estimated financial gain	€688,578 on test set

---

🗂️ Repository Structure

hotel-booking-ml/
│
├── notebook/
│   └── Q3_GMM_clustering.ipynb        # Q3 unsupervised segmentation
│
├── data/
│   └── INNHotelsGroup.csv             # INN Hotels Group dataset (36K bookings)
│
├── report/
│   └── hotel-booking-cancellation-report.pdf  # Full academic report
│
└── README.md

---

🛠️ Stack

Methods: GMM · PCA · BIC/AIC · MAP rule · Unsupervised learning

---

👤 Author

Alexis Mattei — Data Scientist @ Groupe BPCE | MSc Data Science, Paris 1 Panthéon-Sorbonne