Identifying critical nodes via link equations and deep reinforcement learning

Abstract

Identifying an optimal set of nodes that can maximize the spread of influence in a network is a crucial challenge in network science. It has numerous applications such as epidemic control and rumor containment. However, most existing techniques are limited by their high computational costs, making them impractical for graphs with millions of nodes. Moreover, the previous approaches have primarily focused on the structural characteristics of the network while the characteristics of information diffusion are ignored. This paper proposes a deep reinforcement learning framework, DeepELE, to bridge these gaps. DeepELE incorporates a graph embedding technique to represent the graph states and applies a deep reinforcement learning method to learn the policy automatically. Note that we assess the contribution of links to spreading processes and further account for the diffusion-related contribution along with the graph structure information into convolutional neural and the Q network. Extensive experiments on both synthetic and real-world networks validate the efficiency and efficacy of DeepELE. The results demonstrate that DeepELE significantly outperforms the state-of-the-art methods, especially for large-scale networks with millions of nodes.