Dynamic Graph Dismantling with Reinforcement Supervised Learning

There are always some “key” nodes in a big complex network, which can joint the most connected subgraphs. How to identify these nodes, finding a minimum set of nodes to attack for reducing the size of residual network's Largest Connected Component(LCC) to break up the original network, has become a research hotspot. Therefore, a method for determining the “key” nodes based on reinforcement learning framework and supervised learning model is proposed. This algorithm can not only utilize the dynamic exploration ability of reinforcement learning to collect a rich training dataset, but also take advantage of the characteristics that supervised learning is adaptive and has strong generalization ability to possess high efficiency and strong robustness. In order to further improve the algorithm's performance,
$\epsilon$
-greedy mechanism is used to explore more network states. The experiment results show that given the same fraction of removed nodes, our algorithm can make the residual LCC smaller in various networks which is superior to the state-of-the-art algorithms in terms of effectiveness and generalization.