Deep Graph Multi-View Representation Learning With Self-Augmented View Fusion.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Current graph neural network (GNN) methods for multi-view representation learning often concatenate features, potentially losing within-view information and failing to strengthen pivotal views.
• Existing Siamese GNN models may produce uninformative representations due to shared parameters.

Purpose of the Study:

• To propose a novel deep graph auto-encoder for effective multi-view representation learning.
• To address limitations of feature concatenation and parameter sharing in existing GNN approaches.

Main Methods:

• A self-augmented view-weight technique is developed for cross-view fusion to highlight pivotal views.
• Graph neural networks (GNNs) with distinct parameters are used for each view to learn informative representations.
• A neural layer is employed to fit the fusion distribution, enabling end-to-end fusion representation extraction.

Main Results:

• The proposed method demonstrates superior performance in clustering and recognition tasks compared to existing techniques.
• The self-augmented view-weight technique effectively identifies and leverages pivotal views.
• Non-shared parameters in view-specific GNNs lead to more informative representations.

Conclusions:

• The novel deep graph auto-encoder offers an effective solution for multi-view representation learning.
• The proposed approach overcomes key limitations of current GNN-based methods.
• Experimental results validate the model's superior performance on downstream tasks.