AdaGCL+: An Adaptive Subgraph Contrastive Learning Toward Tackling Topological Bias.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Training graph neural networks (GNNs) on large-scale graph data faces scalability challenges.
• Batch sampling methods, used to mitigate scalability issues, introduce topological bias, negatively impacting GNN generalization.
• This bias arises from incomplete subgraph structures, missing node features or edges compared to the complete graph.

Purpose of the Study:

• To propose adaptive subgraph contrastive learning (AdaGCL) to bridge the gap between batch sampling and GNN generalization.
• To develop an enhanced version, AdaGCL+, that automates graph augmentation for improved node embeddings.
• To optimize the augmentation strategy for downstream tasks using a node-centric information bottleneck (Node-IB).

Main Methods:

• AdaGCL augments sampled subgraphs and uses a contrastive loss to learn invariant node embeddings.
• Node-IB is introduced to control the trade-off between similarity and diversity in graph augmentation.
• AdaGCL+ dynamically adjusts graph perturbation parameters to minimize downstream loss, automating augmentation.

Main Results:

• AdaGCL+ demonstrates scalability to graphs with millions of nodes using batch sampling.
• The method consistently outperforms existing approaches on benchmark datasets for node classification accuracy.
• AdaGCL+ shows superior runtime efficiency compared to previous methods.

Conclusions:

• AdaGCL+ effectively addresses the topological bias and generalization issues in large-scale GNN training.
• The automated graph augmentation strategy enhances the performance and efficiency of GNNs.
• AdaGCL+ offers a scalable and effective solution for training GNNs on massive graph datasets.