Learning from Local Walks on Dynamic Graphs with Bandit Feedback
This work identifies a process-agnostic structural condition, based on sliding-window mixing, that ensures the graph's intrinsic walk remains stable for both exploration and navigation and establishes sublinear expected regret.