Seminar by Wenzhuo Zhou

Actor-Critic Graph Neural Networks: A Complete Recipe for Neural Decoding

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. To tackle these issues, we propose a novel Actor-Critic Graph Neural Network (AC-GNN), which is able to integrate information of various orders and provide interpretable results by identifying compact graph structures. In particular, AC-GNN consists of two components: an estimation model for the latent representation of complex relationships under graph topology, and an interpretation model that recover influential nodes, edges, and node features. Theoretically, we establish the generalization error bound for AC-GNN via empirical Rademacher complexity, and showcase its power to represent layer-wise neighborhood mixing. We conduct comprehensive numerical studies using benchmark and simulated datasets to demonstrate the superior performance of AC-GNN in comparison to several state-of-the-art alternatives. Furthermore, we apply AC-GNN to a real-world case study aimed at decoding task-critical information from brain activity data, thereby highlighting its effectiveness in advancing scientific research.