Self-Supervised Node Representation Learning via Node-to-Neighbourhood Alignment.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Artificial Intelligence

Background:

• Self-supervised learning aims to create effective node representations from unlabeled graph data.
• Learning contextual information from graph structure is crucial for informative representations.
• Existing methods face challenges in efficiently capturing neighborhood context and avoiding representation collapse.

Purpose of the Study:

• To develop simple yet effective self-supervised methods for node representation learning.
• To align hidden node representations with their neighborhood context.
• To address memory overheads and representation collapse in contrastive learning frameworks.

Main Methods:

• Proposed a method to maximize mutual information between node and neighborhood representations, acting as graph smoothing.
• Introduced Topology-Aware Positive Sampling (TAPS) for offline positive sample selection in contrastive learning.
• Developed a negative-free solution using a Graph Signal Decorrelation (GSD) constraint to prevent collapse and over-smoothing.

Main Results:

• The node-to-neighborhood alignment theoretically contributes to graph smoothing.
• TAPS enables efficient positive sampling by considering structural dependencies.
• The GSD constraint effectively combats representation collapse and over-smoothing, unifying existing approaches.
• MLP-based encoders with proposed methods show promising node classification results on various datasets.

Conclusions:

• The presented self-supervised methods offer a simple and effective approach to learning node representations.
• The proposed techniques, including TAPS and GSD, enhance graph representation learning efficiency and robustness.
• These methods achieve competitive node classification performance across different scales of graph-structured data.