GNC-Cut: A Hybrid Framework for Community Detection via GNN Embeddings and Classical Clustering

Community detection is essential for understand-ing complicated networks in social media, commerce suggestion, and biological interaction study. Graph Neural Network(GNN)-based techniques are accurate, but their end-to-end designs need GPU resources and confuse the decision process. We provide GNC-Cut, a lightweight, unsupervised framework that decouples nodeembedding, clustering, and boundary refinement. A two-layer GCN encodes structural and attribute signals, Louvain finds initial partitions, and a hybrid cosine-Jaccard criterion sharpens community borders. The modular architecture allows plug-andplay stage replacement and works effectively on ordinary CPUs. Eight real-world networks with 34 to 25000 nodes show that GNCCut decreases average conductance by 10% and matches or exceeds the best baseline in modularity while staying extremely resilient to hyper-parameter variation. These results demonstrate GNC-Cut’s practicality, interpretability, and scalability for community detection in attributed and plain graphs.