Comparative Analysis of Clustering Algorithms: K-Means, DBSCAN, and GMM

📄 Abstract

This repository contains the implementation of our SRC (Student Research Conference, Zayed University) paper on clustering algorithm benchmarking. We present a systematic comparison of three widely used clustering paradigms — K-Means (centroid-based), DBSCAN (density-based), and Gaussian Mixture Models (GMM) (model-based) — across nine benchmark datasets with varying sizes, dimensionalities, and cluster complexities.

All algorithms are evaluated under a unified preprocessing and evaluation pipeline including duplicate removal, outlier filtering, feature standardization, and PCA-based visualization. Performance is assessed using purity, accuracy, macro-precision, macro-recall, macro F1-score, and silhouette score. Results provide a practical decision-making roadmap for clustering algorithm selection based on empirical data characteristics.

📂 Repository Structure

├── data/
│   └── README.md                    # Dataset descriptions and sources
├── notebooks/
│   ├── kmeans_clustering.ipynb      # K-Means implementation and evaluation
│   ├── dbscan_clustering.ipynb      # DBSCAN implementation and evaluation
│   └── gmm_clustering.ipynb         # GMM implementation and evaluation
├── paper/
│   └── Clustering_Comparative_Analysis.pdf   # Published conference paper
├── LICENSE
├── README.md
└── requirements.txt

---

This is the official code for the SRC paper: Comparative Analysis of Clustering Algorithms: K-Means, DBSCAN, and Gaussian Mixture Model Authors: Peter Yacoub, Mohamed Malek Kaouach, Soud Asaad Soud Alhazba, Mohammad Azmi Al-Betar (Ajman University, UAE) Download Paper (PDF)

---

⚙️ Methodology

• Preprocessing: Duplicate removal, Z-score outlier filtering (|z| > 3), feature standardization
• Algorithms: K-Means (RBF), DBSCAN (ε, MinPts tuned via k-distance curve), GMM (EM algorithm)
• Datasets: 9 benchmark datasets — Banknote, Ionosphere, Sonar, Blobs, Varied, Wine, Flame, Glass, Iris
• Evaluation: Purity, Accuracy, Macro-Precision, Macro-Recall, Macro F1-score, Silhouette Score
• Visualization: PCA-based 2D scatter plots for qualitative cluster structure comparison

📊 Results Summary

DBSCAN

Dataset	Purity	Accuracy	F1	Silhouette
Blobs	100.00%	100.00%	1.000	0.81
Flame	100.00%	100.00%	1.000	0.31
Wine	97.46%	97.46%	0.977	0.63
Iris	94.90%	94.90%	0.943	0.67

K-Means

Dataset	Purity	Accuracy	F1	Silhouette
Varied	93.10%	93.10%	0.929	0.63
Wine	94.00%	93.80%	0.933	0.28
Flame	90.90%	67.30%	0.772	0.40

GMM

Dataset	Purity	Accuracy	F1	Silhouette
Blobs	100.00%	100.00%	1.000	0.81
Varied	98.87%	98.87%	0.989	0.60
Ionosphere	72.57%	72.57%	0.621	0.22

DBSCAN achieved the strongest overall robustness on non-convex and noisy datasets. K-Means performed best on compact, well-separated clusters. GMM excelled on elliptical and overlapping structures.

🗃️ Datasets

Nine labeled benchmark datasets were used for evaluation in a fully unsupervised manner:

Dataset	N	Dimensions	Classes
Banknote	1372	4	2
Ionosphere	351	34	2
Sonar	208	60	2
Blobs	1500	2	3
Varied	1500	2	3
Wine	178	13	3
Flame	240	2	2
Glass	214	9	6
Iris	150	4	3

⚠️ Datasets are not included in this repository. See data/README.md for download links.

🚀 Installation

git clone https://github.com/Soudk21/clustering-algorithms-benchmark.git
cd clustering-algorithms-benchmark
pip install -r requirements.txt

📓 Usage

Each notebook is self-contained. Run them independently:

jupyter notebook notebooks/kmeans_clustering.ipynb
jupyter notebook notebooks/dbscan_clustering.ipynb
jupyter notebook notebooks/gmm_clustering.ipynb

🤝 Acknowledgments

• Affiliated with: Artificial Intelligence Research Center (AIRC), Ajman University, UAE
• Submitted to: Student Research Conference (SRC), Zayed University

📜 Citation

If you use this code or findings in your research, please cite:

@inproceedings{yacoub2025clustering,
title={Comparative Analysis of Clustering Algorithms: K-Means, DBSCAN, and Gaussian Mixture Model},
author={Yacoub, Peter and Kaouach, Mohamed Malek and Alhazba, Soud Asaad Soud and Al-Betar, Mohammad Azmi},
booktitle={Proceedings of the Student Research Conference (SRC), Zayed University},
year={2025}
}

License

MIT License. See LICENSE for details.