Abstract:Multiplex network community detection aims to identify densely connected groups of nodes in multi-relational systems. Some network aggregation-based methods rely mainly on the cross-layer occurrence frequency of node pairs across different network layers, while neglecting local topological consistency and intra-layer edge-weight information. As a result, some noisy edges may be assigned excessively high weights, thereby interfering with community partitioning. To address these issues, this paper proposes a multiplex network community detection algorithm based on cross-layer connection frequency and local structural refinement, namely CNPML (Multiplex Network Community Detection via Cross-layer Frequency and Neighborhood Perturbation Fusion). First, the algorithm constructs an initial weighted aggregate graph by adaptively integrating cross-layer frequency, normalized common-neighbor scores, and intra-layer edge-weight information. Second, a structural refinement strategy based on normalized common neighbors is introduced, in which edges are added between highly similar but unconnected node pairs, whereas edges between connected node pairs with low similarity are removed, thereby strengthening potential intra-community connections and suppressing inter-community noisy edges. Finally, the Louvain algorithm is applied to the optimized aggregate graph to obtain the community partition. Experimental results on multiple real-world unweighted and weighted multiplex networks show that CNPML achieves favorable or competitive performance on most datasets and evaluation measures, and can also adapt well to weighted multiplex networks. Ablation experiments and parameter analyses further verify the effectiveness of the individual modules, while runtime experiments demonstrate the scalability of CNPML to large-scale networks.