Identifying Multiple Influential Nodes for Complex Networks Based on Multi-agent Deep Reinforcement Learning

Shengzhou Kong, Langzhou He, Guilian Zhang, Li Tao, Zili Zhang

Abstract: The identification of multiple influential nodes that influence the structure or function of a complex network has attracted much attention in recent years. Distinguished from individual significant nodes, the problem of overlapping spheres of influence among influential nodes becomes a key factor that hinders their identification. Most approaches artificially specify the spacing distance between selected nodes through graph coloring and greedy selection. However, these approaches either fail to find the best combination accurately or have high complexity. Therefore, we propose a novel identification framework, namely multi-agent identification framework (MAIF), which selects multiple influential nodes in a distributed and simultaneous manner. Based on multi-agent deep reinforcement learning, the framework introduce several optimization models and extend to complex networks to solve distributed problems. With sufficient training, MAIF can be applied to real-world problems quickly and effectively, and perform well in large-scale networks. Based on SIR model-based simulations, the effectiveness of MAIF is evaluated and compared with three baseline methods. The experimental results show that MAIF outperforms the baselines on all four real-world networks. This implies using multiple agents to find multiple influential nodes in a distributed manner is an efficient and accurate new way to differentiate from the greedy methods.