Bayesian Adversarial Spectral Clustering with Unknown Cluster Number.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Spectral clustering is widely used in unsupervised learning tasks.
• Deep neural networks have enhanced spectral clustering, but determining the cluster number remains challenging.
• Existing deep spectral clustering methods often struggle with automatic cluster number estimation.

Purpose of the Study:

• To develop a unified Bayesian framework for deep spectral clustering that addresses the challenge of cluster number estimation.
• To integrate spectral clustering, generative adversarial networks (GANs), and low-rank models for improved clustering performance.
• To enable automatic determination of the optimal number of clusters directly from data.

Main Methods:

• A novel Bayesian framework combining spectral clustering, generative adversarial networks, and low-rank models.
• Adaptation of low-rank methods for cluster number estimation within the framework.
• Introduction of a hidden space structure preservation term embedded in the generative process.
• Derivation and implementation of a variational-inference-based optimization and learning procedure.

Main Results:

• The proposed model demonstrates robust cluster number estimation capabilities.
• Experimental results show superior performance compared to existing graph clustering methods.
• The integrated framework effectively learns cluster structures and number from data.

Conclusions:

• The unified Bayesian framework successfully tackles the challenge of automatic cluster number estimation in deep spectral clustering.
• The integration of GANs, low-rank models, and spectral clustering offers a powerful approach for unsupervised learning.
• The method provides a significant advancement over current graph clustering techniques.