Graph Information Bottleneck and AD‑GCL: Enhancing Graph Representation Learning and Robustness

In this talk, guest speaker Li Pan (Assistant Professor, Purdue University) and editor Wu Qiyao (UC San Diego) present an overview of graph representation learning and its connection to graph neural networks (GNNs).

Graph representation learning aims to embed discrete graph structures into continuous vector spaces, enabling downstream tasks such as node classification and graph classification. Traditional methods often capture redundant information, making them sensitive to structural perturbations.

The Graph Information Bottleneck (GIB) framework addresses this by maximizing mutual information between the graph embedding Z and the downstream task label Y while minimizing mutual information between Z and the raw input (features X and adjacency A). This encourages embeddings that retain only the minimal sufficient information needed for the task, improving robustness.

GIB differs from InfoMax, which seeks to preserve as much input information as possible; GIB instead discards irrelevant information, leading to more stable and transferable representations. The optimization involves a trade‑off parameter β and can be implemented by injecting random noise during compression, using re‑parameterization tricks for differentiability.

Experimental results on three benchmark datasets show that GIB outperforms standard GCNs and other baselines under adversarial edge deletions and Gaussian feature noise, demonstrating superior robustness and comparable predictive performance.

To overcome GIB’s reliance on task labels, the authors propose AD‑GCL (Adversarial Graph Contrastive Learning), which integrates GIB ideas into a self‑supervised contrastive learning framework. AD‑GCL learns a learnable edge‑dropping distribution during graph augmentation, forming positive pairs from differently perturbed views of the same graph and negative pairs from different graphs.

AD‑GCL achieves better performance than baseline contrastive methods across multiple datasets and shows promising transfer learning results, outperforming other methods on six out of nine datasets.

The Q&A section discusses practical applications of GIB, the trade‑off between robustness and predictive ability (which is minimal in experiments), and the use of re‑parameterization for Gaussian sampling.

Overall, the talk highlights how information‑theoretic principles can guide the design of more robust and label‑efficient graph learning algorithms.