Learning the Heuristics of Hub Identification over Brain Networks

A plethora of neuroscience studies find that connector hub nodes play a key role in regulating multiple modules and supporting brain functions such as consciousness and cognition, due to its critical topological location in the network. Current approaches mainly rely on the 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, which often results in less accurate hub identification results. Although graph theory allows us to find connector hubs with a greater mathematical insight, the large scale of brain network often compromises the well-defined optimization into a local and sub-optimal solution. To overcome these limitations, we propose a joint graph embedding and hub identification solution 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 the current state-of-the-art hub identification methods. We have achieved more reliable and replicable hub identification results on both simulated and real brain network data, suggesting the high applicability to various network analysis studies in neuroscience and neuroimaging fields.