DRL-DC-DQN: A Deep Reinforcement Learning Approach to Decentralized Community Detection

Abstract

Decentralized community detection remains a fundamental challenge in network analysis, requiring methods that operate without central coordination using only local information. This paper introduces DRL-DC-DQN (Decentralized Reinforcement Learning for Community Detection with Deep Q-Network), a novel multi-agent reinforcement learning framework where each network node functions as an autonomous agent. Agents decide whether to remain in their current community or transition to a neighboring community based on local observations and neighbor memberships, guided by a neural network approximating the Q-function. Learning is driven by a locally defined reward function that evaluates community cohesion, incentivizing agents to form structurally uniform communities. The deep reinforcement learning paradigm inherently addresses scalability concerns while improving decision quality in large-scale, complex networks. Experimental evaluations on real-world datasets demonstrate that DRL-DC-DQN achieves superior efficiency and robustness compared to state-of-the-art decentralized approaches, establishing its effectiveness for distributed community detection tasks.