Multi-sample dual-decoder graph autoencoder.

Area of Science:

• Graph representation learning
• Self-supervised learning
• Machine learning

Background:

• Self-supervised learning excels in graph tasks, but contrastive methods face limitations with data augmentation and complex training.
• Generative approaches like graph autoencoders (GAEs) offer an alternative, avoiding these dependencies.
• Existing GAEs often focus on either graph topology or node features, neglecting their complementary information.

Purpose of the Study:

• To propose a novel generative self-supervised graph representation learning methodology.
• To address limitations of existing methods by reconstructing both graph structure and node features.
• To enhance generalization through a multi-sample learning strategy.

Main Methods:

• Introduced the Multi-View Dual-decoder Graph Autoencoder (MDGA).
• Employed a multi-sample graph learning strategy for improved generalization.
• Implemented dual decoders for reconstructing graph topology (traditional GAE) and node attributes (masked feature reconstruction).

Main Results:

• MDGA demonstrated superior performance across five public benchmark datasets.
• Outperformed state-of-the-art methods in both node classification and link prediction tasks.
• The dual-decoder approach effectively leveraged complementary information from graph structure and node features.

Conclusions:

• MDGA offers an effective generative self-supervised approach for graph representation learning.
• The model successfully integrates topological and feature reconstruction for enhanced performance.
• This methodology provides a promising direction for advancing graph-based machine learning tasks.