Differentiable Hierarchical Optimal Transport for Robust Multi-View Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Traditional multi-view learning assumes aligned samples and shared latent space distributions, limiting practical applications.
• These assumptions are often violated in real-world scenarios, necessitating more robust methods.

Purpose of the Study:

• To propose a novel differentiable hierarchical optimal transport (DHOT) method for multi-view learning.
• To overcome the limitations of traditional multi-view learning by handling unaligned data and distributional discrepancies.

Main Methods:

• Calculates sliced Wasserstein distance between latent distributions of unaligned multi-view data.
• Employs entropic optimal transport to identify view clustering structures.
• Defines a hierarchical optimal transport distance as the objective function for bi-level optimization, treating entropic optimal transport as a differentiable operator.

Main Results:

• The DHOT method demonstrates superiority over traditional alternating optimization strategies.
• Achieves performance comparable to state-of-the-art methods on synthetic and real-world tasks, particularly with unaligned data.
• Applicable to both unsupervised and semi-supervised learning settings.

Conclusions:

• DHOT effectively mitigates dependency on alignment and distribution assumptions in multi-view learning.
• The bi-level optimization strategy with differentiable optimal transport enhances model training.
• DHOT offers a robust and versatile approach for challenging multi-view learning problems.