DISCOVERING UNPRECEDENTED HEURISTICS FOR HUB IDENTIFICATION BY JOINT GRAPH EMBEDDING AND REINFORCEMENT LEARNING

A plethora of neuroscience studies find that connector hub nodes play a key role in regulating multiple modules and supporting brain functions, due to its critical topological location in the network. Current methods mainly rely on hand-crafted attributes (aka. graph embedding at each node) from the domain knowledge of network neuroscience such as high connectivity degree to identify connector hub nodes. However, simple ranking heuristic based on the pre-defined attributes has limited power to characterize the complex network topology, w Inch often results in less accurate hub identification. Although graph theory allows us to find connector hubs. the large scale of brain network often compromises the well-defined optimization into a local and sub-optimal solution. To overcome, we propose a joint graph embedding and hub identification in a reinforcement learning framework to discover the unprecedented heuristics from the existing knowledge of network neuroscience and graph theory, which allows us to outperform state-of-the-art hub identification methods.