Multi-View Subspace Clustering via Structured Multi-Pathway Network.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Deep multi-view clustering (MVC) shows promise but often uses single-pathway networks.
• Existing methods struggle to capture comprehensive complementary and multilevel features across views.

Purpose of the Study:

• To propose a novel deep structured multi-pathway network (SMpNet) for enhanced multi-view subspace clustering.
• To address limitations in feature extraction by exploring complementary information and multilevel features.

Main Methods:

• Developed SMpNet using structured multi-pathway convolutional neural networks for layer-wise subspace representation learning.
• Integrated low-level and high-level structured features via a common connection matrix.
• Applied a low-rank constraint on the connection matrix to reduce noise and emphasize consensus information.

Main Results:

• Demonstrated the effectiveness of SMpNet on five public datasets.
• Achieved superior performance compared to several state-of-the-art deep MVC methods.
• Validated the ability to explore comprehensive complementary structures among multiple views.

Conclusions:

• SMpNet successfully integrates multilevel features and complementary information for improved multi-view subspace clustering.
• The proposed method offers a significant advancement in deep multi-view learning.
• The layer-wise feature integration and low-rank constraint contribute to robust and effective clustering.