sonsc-isi 0.1

Improved SONSC+ clustering evaluation using ISI (Improved SONSC Index)

SONSC+ (ISI): Optimal Number of Clusters Selection Using ISI Index

A Python implementation of the improved SONSC+ algorithm utilizing the Improved Separation Index (ISI) for accurate clustering evaluation and optimal selection of clusters without requiring labeled data.

📌 Installation

Install via pip:

pip install sonsc-isi

🚀 Features

Automatic selection of the optimal number of clusters.

High interpretability using Cohesion, Separation, and ISI.

Compatible with various datasets including MNIST, CIFAR-10, Iris, etc.

Generates detailed visualizations for easy interpretation.

🔖 Usage Example

Here's how to use sonsc_isi in your project:

from sonsc_isi import sonsc_isi from sklearn.datasets import load_digits from sklearn.preprocessing import StandardScaler import matplotlib.pyplot as plt

X, _ = load_digits(return_X_y=True) X = StandardScaler().fit_transform(X)

k_values = range(2, 15) results_df = sonsc_isi(X, k_values=k_values)

print(results_df)

Visualizing results

plt.figure(figsize=(10, 5)) plt.plot(results_df["k"], results_df["ISI"], marker='o', label='ISI Score', color='purple') plt.axvline(results_df.loc[results_df["ISI"].idxmax(), "k"], linestyle='--', color='red', label='Optimal k') plt.title("ISI Scores for Different Numbers of Clusters") plt.xlabel("Number of Clusters (k)") plt.ylabel("ISI Score") plt.legend() plt.grid() plt.show()

📚 Supported Datasets

MNIST

CIFAR-10

Iris

Digits

Any custom numerical dataset

🖥 Example Notebook

See examples/example_sonsc_isi.ipynb for detailed usage examples on MNIST and CIFAR-10 datasets.

📖 Reference

The algorithm is inspired by and improved upon the work described in:

Cosine Clustering Index (CCI), SN Computer Science

@article{CCI2024, author = {Jahanian, Mojtaba and Others}, title = {Cosine Clustering Index (CCI)}, journal = {SN Computer Science}, year = {2024}, doi = {10.1007/s42979-024-02970-7} }

⚙️ Dependencies:

numpy

pandas

matplotlib

scikit-learn

tensorflow

📄 License:

MIT © Your Name