Deep Supervised Multi-View Learning With Graph Priors.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view representation learning aims to find a common latent space for data from different sources.
• Existing methods often struggle with preserving both discriminative information and the intrinsic structure of individual views.
• Scalability to large datasets remains a challenge in multi-view learning.

Purpose of the Study:

• To develop a novel supervised multi-view representation learning method.
• To project multiple data views into a shared latent space.
• To preserve the discrimination and intrinsic structure of each view within the common space.

Main Methods:

• Constructing an apriori discriminant similarity graph based on labels and pairwise relationships.
• Utilizing view-specific networks to map inputs to common representations.
• Enforcing constraints for graph consistency, discrimination, and cross-view invariance.
• Implementing a sampling strategy for efficient approximation of the similarity structure.

Main Results:

• The proposed method successfully preserves intrinsic structure and discrimination in the latent common space.
• The sampling strategy improves space complexity and enables handling of large-scale multi-view datasets.
• Experimental results on five datasets demonstrate superior performance compared to 18 state-of-the-art methods.

Conclusions:

• The novel method effectively achieves supervised multi-view representation learning.
• The approach offers a scalable and efficient solution for complex multi-view data.
• The findings suggest significant advancements in the field of representation learning.