Academic

In recent years, graphs have become one of the most powerful abstractions for complex data, including brain networks, knowledge graphs, purchasing behavior, as well as disease pathways. Among many graph representation learning approaches, Graph Neural Networks (GNNs) have achieved promising performance in a variety of graph-focused tasks. Despite their success, existing GNNs suffer from two significant limitations that hinder their broader applications in scientific discovery: a lack of interpretability in results due to their black-box nature, and an inability to learn representations of varying order information with statistical guarantees.

In this talk, I will present a novel Actor-Critic Graph Neural Network (AC-GNN), which is able to integrate information of various orders under graph topology and provide interpretable results by identifying compact subgraph structures. Statistically, we establish the generalization error bound for AC-GNN via empirical Rademacher complexity, and showcase its power to represent layer-wise neighborhood mixing. Comprehensive numerical experiments using benchmark and synthetic datasets are conducted. Interestingly, we use AC-GNN to decode the behavior-oriented information from neuronal activity signals and confirm several important conjectures in neurobiology, thereby highlighting its effectiveness in advancing scientific research.