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

Data in the real world often takes the form of graphs—social networks, biological systems, knowledge graphs—and graph representation learning has become a prominent research direction. This talk examines graph representation learning from an information‑theoretic perspective, focusing on three topics: an introduction to graph representation learning, the Graph Information Bottleneck (GIB), and AD‑GCL (Graph Contrastive Learning based on GIB).

Graph representation learning aims to map discrete graph structures into vector spaces for downstream tasks. Using message‑passing neural networks (GNNs), node features are iteratively updated, and graph‑level embeddings are obtained via pooling. This provides a flexible, powerful alternative to classical graph kernels.

GIB addresses the sensitivity of GNNs to structural perturbations by seeking representations that capture minimal sufficient information for the downstream task while discarding redundant information from the input graph. Its objective maximizes the mutual information I(Y;Z) between the representation Z and task labels Y, and minimizes I(D;Z) between Z and the raw data D, contrasting with the InfoMax principle that maximizes I(D;Z).

The GIB algorithm proceeds in a compression‑then‑prediction fashion. Randomness (e.g., edge sampling) is injected during compression to reduce I(D;Z); then a standard GNN message‑passing step updates node features on the perturbed graph. The whole process is trained via variational optimization, allowing the model to discard task‑irrelevant information while preserving task‑relevant cues.

Experimental results on three benchmark datasets, evaluated under adversarial attacks (edge deletion, Gaussian noise), show that GIB consistently outperforms vanilla GCN and other baselines, demonstrating superior robustness and stability without sacrificing accuracy on clean data.

AD‑GCL extends GIB to a self‑supervised setting where task labels are unavailable. It employs adversarial graph augmentation: a learnable edge‑dropping probability distribution generates two augmented views of the same graph, which serve as a positive pair in a contrastive loss. Negative pairs are formed from different graphs. This framework eliminates the need for labeled data while still encouraging representations that retain sufficient information.

AD‑GCL experiments compare the method with baseline contrastive approaches and random edge‑dropping strategies across multiple datasets. AD‑GCL achieves the best performance in most cases and also shows strong transfer learning ability, ranking first on six out of nine evaluated datasets.

The Q&A section highlights practical applications of GIB (node classification, molecule‑level graph classification), discusses the trade‑off between robustness and predictive power (found to be negligible in experiments), and explains the use of the re‑parameterization trick to keep Gaussian sampling differentiable.

The talk concludes with acknowledgments and a reminder to like, share, and follow the DataFunTalk platform for more AI and big‑data content.