This project explores and compares the effectiveness of ensemble clustering methods versus traditional base clustering algorithms for customer segmentation. The analysis is performed on three real-world datasets, and the results are thoroughly evaluated using multiple clustering metrics and stability analyses. The project is accompanied by a scientific paper that details the methodology, results, and insights.

final_submission/
├── code/         # All Python scripts and modules
├── dataset/      # Datasets used for analysis
├── plots/        # Generated plots and visualizations (output)
├── results/      # Output metrics and results (generated at runtime)
├── doc/          # Documentation and paper PDF
├── .gitignore
└── README.md

• Mall_Customers.csv
• customer_personality.csv
• wholesale_customers.csv

All datasets are located in the dataset/ folder.

• Base Algorithms:
  • K-Means
  • Agglomerative Clustering
  • DBSCAN
  • HDBSCAN
• Ensemble Methods:
  • CSPA (Cluster-based Similarity Partitioning Algorithm)
  • HGPA (HyperGraph Partitioning Algorithm)
  • MCLA (Meta-CLustering Algorithm)

• Silhouette Score
• Davies-Bouldin Index
• Calinski-Harabasz Index
• Average Normalized Score
• Stability Analysis (Bootstrap ARI, Variation of Information)
• Noise Stability

1. Set up the environment:

• Recommended: Use the provided conda environment (e.g. conda activate ai).
• Required packages: scikit-learn, matplotlib, seaborn, pandas, numpy, hdbscan, etc.

2. Install dependencies:

python3 -m pip install -r code/requirements.txt

3. Run the main script (from this final_submission/ folder):

python3 code/main.py --dataset mall_customers
python3 code/main.py --dataset customer_personality
python3 code/main.py --dataset wholesale_customers

Outputs:

• Plots: plots/
• Metrics CSV: results/metrics_<dataset>.csv

Interpretations and summaries are printed to console.

A detailed scientific paper describing the methodology, experiments, and findings is available in doc/main.pdf (PDF).

• Literature review
• Methodological details
• Algorithmic explanations
• Experimental results
• Discussion and conclusions

• All code is parameterized and can be run on any of the provided datasets without modification.
• The project structure and CLI interface ensure easy testing and extension.
• The .gitignore excludes only temporary, cache, and output files (e.g., plots/, Python cache, LaTeX build files).

Francesco Albano [fra2404].

This project is for academic purposes. For reuse or distribution, please contact the author.