A New Q-Learning-Based Community Detection Algorithm

Abstract

Community detection is the core problem in complex network analysis. Although traditional community detection methods have made some progress, they are often too sensitive to initial conditions and local optimal solutions, and have high computational complexity. To address these difficulties, this paper proposes a new community detection method based on $Q$-learning by defining reward function and new metrics. The new reward function combines modularity and modularity density to ensure the tightness of the community structure and the rationality of the whole network partition, while the new metric evaluates the centrality of nodes is used to improve the accuracy of candidate node selection and reduce invalid searches. Meanwhile,  $\epsilon$-greedy strategy and $Q$-table update are also applied such that the community partition is more fine-grained, and the accuracy is still high in noisy datasets. Moreover, a method of finding the optimal parameter setting is also  given. Experiments on a variety of datasets show that the proposed method performs well  and is robust. The proposed method provides an effective solution for community detection in complex networks, and is suitable for different network datasets.

Keywords: Community detection\sep node similarity\sep algorithm robustness \sep Q-learning