Multi-View Diffusion Process for Spectral Clustering and Image Retrieval.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view graph learning constructs unified affinity graphs from heterogeneous data.
• Existing methods often rely on initial graph quality or latent subspace assumptions.
• Adaptive fusion of multiple data representations is crucial for robust learning.

Purpose of the Study:

• To develop a novel multi-view graph learning framework combining weight and graph learning.
• To propose a fusion-and-diffusion strategy for adaptive graph construction.
• To enhance learning by leveraging high-order contextual information and manifold awareness.

Main Methods:

• An alternating optimization framework integrates weight learning and graph learning.
• A fusion-and-diffusion strategy uses unsupervised graph smoothness for adaptive graph merging.
• A novel multi-view diffusion process propagates affinities on tensor product graphs.

Main Results:

• The approach effectively identifies and fuses consistent information across multiple views.
• It outperforms state-of-the-art methods in image retrieval and clustering on 13 out of 16 datasets.
• The method is robust and not limited by initial graph quality or common subspace assumptions.

Conclusions:

• The proposed unified framework offers a powerful approach to multi-view graph learning.
• Adaptive fusion and manifold-aware diffusion significantly improve performance.
• The method demonstrates broad applicability and superior results in real-world tasks.