Redundancy is Not What You Need: An Embedding Fusion Graph Auto-Encoder for Self-Supervised Graph Representation...

Area of Science:

• Graph machine learning
• Data mining and analysis
• Network science

Background:

• Attribute graphs are vital for graph communities but suffer from redundancy and noise.
• These issues distort data, compromising attribute graph learning accuracy and reliability.
• Overfitting and underfitting can occur due to irrelevant or noisy attributes and structural features.

Purpose of the Study:

• To propose an embedded fusion graph auto-encoder (EFGAE) framework for self-supervised learning (SSL).
• To address redundancy and noise in attribute graphs for improved learning.
• To enhance the accuracy and robustness of attribute graph learning.

Main Methods:

• The EFGAE framework utilizes multitask learning to fuse node features across tasks.
• It involves a pretraining phase using adversarial contrastive learning within a graph auto-encoder (GAE).
• A downstream task learning phase employs an adaptive graph convolutional network (AGCN) for GNN classifiers.

Main Results:

• The EFGAE framework effectively reduces redundancy and noise in attribute graphs.
• Experimental results show superior performance compared to state-of-the-art (SOTA) methods.
• The approach demonstrates enhanced accuracy, generalization ability, and robustness.

Conclusions:

• The proposed EFGAE framework offers a robust solution for attribute graph learning.
• Self-supervised learning combined with feature fusion effectively mitigates data imperfections.
• This method significantly advances graph representation learning for complex network analysis.