Q -learning-based community detection algorithm

Community detection is a central problem in complex network analysis, yet conventional algorithms often suffer from sensitivity to initialization, entrapment in local optima, and high computational costs. We propose a community detection framework based on multiagent reinforcement learning that integrates a reward function balancing intracommunity compactness and intercommunity separateness, an initialization strategy guided by node importance, and node embeddings trained via DeepWalk. Each community is assigned an independent detection agent that allocates nodes through a deep Q-learning network, enabling adaptive partitioning. The use of node importance and embedding distance improves candidate selection efficiency, while the ε-greedy strategy and target network updates enhance global exploration and stability. Extensive experiments on multiple datasets demonstrate that the proposed method consistently outperforms baseline approaches, achieving more accurate and scalable community detection, particularly in large-scale networks.