Nonparametric Clustering-Guided Cross-View Contrastive Learning for Partially View-Aligned Representation Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view representation learning is crucial with increasing multi-view data availability.
• Collecting strictly view-aligned data is expensive, making Partially View-aligned Representation Learning (PVRL) practical.
• Existing PVRL methods struggle with incomplete correspondences and false negative pairs (FNP).

Purpose of the Study:

• To address limitations in existing PVRL methods, particularly regarding false negative pairs (FNP).
• To propose a novel method for robust and effective Partially View-aligned Representation Learning.
• To improve downstream tasks like clustering using enhanced multi-view representations.

Main Methods:

• Proposed Nonparametric Clustering-guided Cross-view Contrastive Learning (NC3L) for PVRL.
• Estimated similarity matrix using marginal cross-view contrastive loss to approximate supervised contrastive learning (CL).
• Developed Deep Variational Nonparametric Clustering (DeepVNC) to discover FNP and construct cluster-level similarity.
• Established theoretical foundation by analyzing error bounds of the loss function.

Main Results:

• NC3L demonstrated superiority over state-of-the-art methods on four benchmark datasets.
• The proposed DeepVNC effectively identifies and handles false negative pairs.
• The theoretical analysis provides a foundation for the method's effectiveness.
• Improved robustness and performance of the contrastive learning method via a reparameterization trick.

Conclusions:

• The proposed NC3L method offers a significant advancement in Partially View-aligned Representation Learning.
• NC3L effectively overcomes limitations of existing methods in handling data incompleteness and false negative pairs.
• The approach provides a theoretically grounded and empirically validated solution for multi-view representation learning.