Simple yet effective heuristic community detection with graph convolution network

In recent years, graph neural network-based community detection methods have integrated local structure and node attributes, incorporating various optimization strategies with notable progress. However, most current algorithms require predefining the number of communities, introducing human bias, and rely on contrastive objectives or data augmentation, leading to extra hyperparameters and complexity. To address these issues without sacrificing detection quality, we propose an adaptive community detection framework that eliminates contrastive learning and the need for pre-specified community numbers, simplifying training and reducing prior dependency. First, the adaptive detection method is introduced to ensure the identification of high-quality structural communities as reliable global references. Then, a novel mechanism for modeling node-community relationships is proposed, integrating global structure, local structure, and attribute information into a unified space. Finally, a reconstructed soft modularity loss is applied to optimize node-community relationships end-to-end, enhancing community structure without data augmentation or contrastive learning. The proposed approach is efficient to train and computationally lightweight, demonstrating superior detection efficiency and competitive accuracy across multiple graph datasets compared to traditional and recent deep learning methods. The code is available at https://github.com/wuanghoong/Less-is-More.git .