Multiclass semisupervised learning based upon kernel spectral clustering.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Semisupervised learning addresses challenges with limited labeled data in machine learning.
• Kernel Spectral Clustering (KSC) is a powerful technique for data clustering.
• Efficiently leveraging unlabeled data is crucial for improving model performance.

Purpose of the Study:

• To propose a multiclass semisupervised learning algorithm based on Kernel Spectral Clustering (KSC).
• To enhance label propagation to large unlabeled datasets using regularization.
• To optimize the embedding dimension for semisupervised clustering with numerous clusters.

Main Methods:

• A regularized Kernel Spectral Clustering (KSC) model is formulated for semisupervised settings.
• A one-versus-all strategy is employed to estimate class memberships for both labeled and unlabeled data.
• Label propagation is achieved by adding regularization terms to the KSC cost function, solved via a linear system in the dual.
• An optimal embedding dimension is determined for effective semisupervised clustering.

Main Results:

• The proposed regularized KSC effectively propagates labels to unlabeled data points.
• The method demonstrates robust performance in multiclass semisupervised clustering scenarios.
• The optimal embedding dimension selection improves clustering accuracy, especially with many clusters.

Conclusions:

• The developed algorithm offers an effective approach for multiclass semisupervised learning.
• Regularization is key to successful label propagation in KSC-based semisupervised clustering.
• The method provides a valuable tool for scenarios with abundant unlabeled data and numerous classes.