Graph Information Bottleneck

Graph Information Bottleneck (GIB) focuses on learning compact and informative graph representations by minimizing redundant information while preserving task-relevant details. Current research emphasizes developing GIB-based methods for improved robustness against adversarial attacks, enhanced explainability of Graph Neural Networks (GNNs), and efficient communication in large graphs, often incorporating techniques like vector quantization and self-supervised learning within GNN architectures. This work is significant because it addresses limitations in GNNs, such as over-squashing and heterophily, leading to more accurate, robust, and interpretable models with applications in diverse fields like medical imaging, remote sensing, and multi-agent systems.