Contrastive Multi-View Kernel Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Kernel methods are established in multi-view learning, defining Hilbert spaces for linear separation.
• Existing algorithms compute kernels independently per view, neglecting cross-view complementary information and potentially leading to suboptimal kernel choices.

Purpose of the Study:

• To propose a novel kernel function, the Contrastive Multi-view Kernel, utilizing the contrastive learning framework.
• To develop a contrastive multi-view clustering framework compatible with existing kernel theory and applications.

Main Methods:

• The Contrastive Multi-view Kernel implicitly embeds multiple views into a joint semantic space, promoting similarity within views and diversity across them.
• The proposed framework is instantiated with multiple kernel k-means for clustering tasks.

Main Results:

• The effectiveness of the Contrastive Multi-view Kernel was validated through extensive empirical studies.
• The contrastive multi-view clustering framework achieved promising performance, demonstrating the benefits of the novel kernel approach.

Conclusions:

• This work pioneers the exploration of kernel generation within a multi-view setting using contrastive learning.
• The proposed Contrastive Multi-view Kernel offers a new paradigm for multi-view kernel learning, enhancing performance by integrating cross-view information.