Liang Zhao is Winship Distinguished Associate Professor of Computer Science at Emory University, with 200+ papers across prestigious venues including KDD, NeurIPS, ICLR, TPAMI, and PIEEE. His honors include the 2025 SIGKDD Test of Time Award for the EMBERS civil-unrest forecasting paper, ICDM Best Paper Awards (2019, 2022), and an NSF CAREER Award. Industry recognitions include the Amazon Research Award, Meta Research Award, and Cisco Faculty Research Award. He earned his Ph.D. from Virginia Tech as Outstanding Doctoral Student (2017) and serves as a Computing Innovation Fellows mentor. His group develops graph/geo-spatiotemporal machine learning—especially graph neural networks and optimization—for spatio/temporal/textual complex networks.
Presentation Title:
Unified GNNs for Dynamic Graph Inference and Optimization: Toward Power-Grid Foundation Models
Presentation Abstract:
Energy systems such as power grids hinge on dynamic graph optimization and graph inverse problems: localizing faults and attacks, estimating hidden states, and choosing interventions under constraints. I will present a unified graph-neural framework that replaces hand-crafted heuristics with principled, learnable solvers, spanning: source localization (SL-VAE, invertible diffusion, cross-network); influence maximization with guarantees (DeepIM, MIM-Reasoner); neural network interdiction; path-centric network tomography; and propagation-tree identification. Across these tasks, we meld generative inference with structure-aware GNNs to recover latent causes, forecast spread, and compute robust control actions. I will close by our recent progress on unifying multiple power-grid optimization problems toward foundation models that fuse physics, optimization, and graph topology.