Researchers from IMT Atlantique and Orange Labs in France have introduced three novel approaches to enhance unsupervised clustering algorithms using labeled data.

The researchers involved are Colin Troisemaine, Alexandre Reiffers-Masson, Stephane Gosselin, Vincent Lemaire, and Sandrine Vaton.

The three approaches are named NCD K-means, NCD Spectral Clustering, and Projection-Based NCD.

These methods aim to optimize hyperparameters without prior knowledge of novel classes.

A simple deep NCD model has also been introduced that accurately estimates the number of novel classes in tabular data.

By adapting k-means and Spectral Clustering algorithms to leverage known class information, the study provides a robust method called PBN (Projection-Based NCD).

The PBN method addresses the challenges of NCD without making unrealistic assumptions.

The code for the proposed methods is available on GitHub for further exploration and insights.