Image Clustering With Transition Probabilities Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-view clustering for image data shows high performance but often lacks interpretability.
• Existing methods may not fully exploit the complementary distributions across different data views.

Purpose of the Study:

• To develop a novel multi-view clustering method that enhances interpretability and utilizes inter-view complementarity.
• To introduce a probabilistic framework for robust clustering of image data.

Main Methods:

• Developed an anchor graph factorization model based on transition probabilities.
• Learned transition probability matrices for samples-to-clusters and anchor points-to-clusters, acting as soft labels.
• Applied Schatten p-norm regularization to align cluster information across multiple views.

Main Results:

• Achieved one-step label acquisition with a sound probabilistic interpretation.
• Effectively mined complementary information among views by aligning labels.
• Demonstrated effectiveness on both small-scale and large-scale image datasets.

Conclusions:

• The proposed Multi-View Clustering with Transition Probabilities Learning (MVC-TPL) method offers improved interpretability and clustering performance.
• The transition probability learning and Schatten p-norm regularization effectively capture inter-view dependencies.