Tourism Package Purchase Prediction

This project builds a supervised learning pipeline to predict whether a customer is likely to purchase a newly introduced travel package. The model is designed to assist the "Visit With Us" tourism company in strategically targeting high-potential customers based on demographic and behavioral features. The data was provided by MIT IDSS within my enrollment in the course "DATA SCIENCE AND MACHINE LEARNING: MAKING DATA-DRIVEN DECISIONS."

---

Project Structure

tourism-purchase-prediction/
├── data/
│   └── README.md                        # Context for absent data
├── notebooks/
│   └── tourism-purchase-prediction.ipynb  # Full EDA and model walkthrough
├── docs/
│   └── index.html                       # GitHub Pages visual summary
├── scripts/
│   ├── model_pipeline.py               # Clean modeling pipeline
│   └── utils.py                        # Helper plotting functions
├── report/
│   └── summary.md                      # EDA + insights + modeling reflection
├── results/
│   └── final_metrics.txt              # Final model evaluation report
├── images/
│   ├── age_distribution.png
│   ├── income_distribution.png
│   ├── trips_distribution.png
│   ├── marital_status_vs_target.png
│   ├── productpitched_vs_target.png
│   ├── passport_vs_target.png
│   ├── designation_vs_target.png
│   ├── feature_importance_dt.png
│   ├── feature_importance_rf.png
│   ├── pr_curve_logreg.png
│   ├── pr_curve_svm_rbf.png
│   └── pr_curve_dt.png
├── README.md                           # You're here
└── requirements.txt                    # Dependencies

---

Problem Overview

The task is a binary classification problem:

• Target: ProdTaken (1 = Purchased, 0 = Did not purchase)
• The dataset includes 20+ features: age, income, occupation, passport status, marital status, pitch satisfaction, etc.

---

Exploratory Analysis & Business Insights

The EDA revealed critical patterns:

• Single customers are more likely to convert than married ones.
• Executives convert more frequently than upper-level designations.
• Customers with a passport and those who self-inquire show higher purchase rates.
• The basic and standard packages had higher conversion rates than deluxe options.

See report/summary.md for full findings and interpretation.

---

Modeling Techniques

The following models were developed and evaluated:

• Logistic Regression
• Support Vector Machine (Linear + RBF kernels)
• Decision Tree (baseline + tuned)
• Random Forest (baseline)

Hyperparameter tuning was done via GridSearchCV, focusing on maximizing recall due to business needs (avoiding loss of potential customers).

---

Final Model Performance

The final model used an SVM with RBF kernel, with an optimal threshold determined via PR curve.

See results/final_metrics.txt for full details.

Best Model: SVM (RBF) with threshold 0.17

• Recall: 0.69
• Precision: 0.48
• F1 Score: 0.56
• RMSE: 0.89

Visual metrics:

• images/pr_curve_svm_rbf.png
• images/feature_importance_rf.png

---

How to Run

1. Install Dependencies

pip install -r requirements.txt

2. Run the Model Pipeline

python scripts/model_pipeline.py

Resources

Dataset from MIT IDSS: "Data Science and Machine Learning: Making Data-Driven Decisions"

Original assignment: Classification and Hypothesis Testing Practice Project

Author

Eliana Gabriela Matos Polanco Email: elianagmpolanco@gmail.com GitHub: @elianagm

Please see the project report and business conclusions in report/summary.md, or in the GitHub page for this project