Local-Global Geometric Information and View Complementarity Introduced Multiview Metric Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Geometry provides foundational insights into spatial relationships and object locations, crucial for classification tasks.
• A geometric perspective offers a novel approach to understanding sample information and improving classification accuracy.

Purpose of the Study:

• To propose a novel multiview metric learning method, GIVCMML, that effectively utilizes local-global geometric information and view complementarity.
• To enhance the separability of samples in a learned metric space by preserving geometric relations.

Main Methods:

• Introduced global geometrical constraints within the maximum margin criterion to maximize distances between class centers.
• Constructed an adjacency matrix incorporating sample label information to explore local geometric information.
• Maximized correlation between views in the metric space to exploit complementary information and enable adaptive learning.

Main Results:

• The proposed GIVCMML method effectively exploits geometric information from multiview samples.
• The learned metric space retains geometric relations, leading to improved sample separability.
• Experimental results on real-world datasets show GIVCMML achieves competitive performance against existing multiview metric learning (MvML) methods.

Conclusions:

• GIVCMML successfully integrates local and global geometric information with view complementarity for enhanced classification.
• The method demonstrates superior performance in retaining geometric structures and improving sample discrimination.
• GIVCMML offers a promising advancement in multiview metric learning, outperforming traditional MvML approaches.