GSCpy 0.1.1

Generalized Spectral Clustering in Python

This is a package for performing Spectral Clustering.

It works a near-fully unsupervised way : the only required information is the number of clusters .

The clustering is done using the Generalized Spectral Clustering (GSC) framework developped by Jonckheere et al. in arXiv:2203.03221. It has been shown experimentally that this framework regularly outperforms classical spectral clustering for synthetic and real datasets.

Classical spectral clustering can also be performed by tweaking the parameters, as the clustering algorithm used is fully customizable.

Interacting with the package is done trough the GSC class, representing a GSC model. To use :

• Create a GSC object with the parameters of your choice
• Cluster your data using the fit method of the class
• Retrieve the clustering using the labels attribute
• Get more information on the clustering by using the available instance attributes (cluster centers, eigenvalues of the graph laplacian, adjacency matrix, Calinski-Harabasz index)
• Evaluate the performance of the clustering using the nmi method.

To help you manage your datasets, GSCpy includes a file manager allowing to easily load and save datasets with their labels.The package also includes an interactive 2D dataset builder, powered by matplotlib.

GSCpy is entirely written in Python and requires the following libraries to run correctly :

• NumPy
• Matplotlib
• SciPy
• Scikit-learn

This project was carried out as part of an internship at LAAS-CNRS, Toulouse.