Beyond local aggregation: Global graph contrastive learning for multi-view fusion.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view fusion is crucial for integrating heterogeneous data sources.
• Unsupervised graph neural network-based multi-view learning faces challenges in graph construction, alignment, and information exploitation.

Purpose of the Study:

• To propose the Global Graph Contrastive learning for Multi-view fusion (G²CM) algorithm.
• To address the key challenges in unsupervised multi-view learning using graph neural networks.

Main Methods:

• G²CM integrates global topology with view-specific weighted edges for reliable graph construction.
• A contrastive learning framework with carefully designed positive and negative pairs enhances cross-view alignment.
• Distance-aware scaling in the loss function improves the exploitation of structural information.

Main Results:

• G²CM achieves state-of-the-art performance across six benchmark multi-view datasets.
• The method demonstrates effectiveness on diverse data types.
• Experimental results validate the proposed approach for multi-view fusion.

Conclusions:

• G²CM effectively addresses limitations in unsupervised multi-view learning.
• The algorithm enhances representation learning by integrating global and local structural information.
• The proposed method offers a robust solution for multi-view fusion tasks.