Similarity-smooth graph contrastive learning for community detection in adaptive oscillatory networks

Community detection in adaptive oscillatory networks is crucial for understanding network evolution, synchronization dynamics, and the underlying mechanisms of adaptive behavior. However, most existing methods focus on static networks and fail to capture the inherently dynamic nature of adaptive oscillatory networks. Although several approaches have been developed to track the temporal dynamics, they often rely solely on network topology and overlook high-dimensional node attributes. Additionally, these methods struggle to effectively integrate global structural patterns with local node interactions, leading to suboptimal performance. To address these challenges, we propose a new method, named Similarity-Smooth Graph Contrastive Learning (SSGCL), for community detection in adaptive oscillatory networks. Firstly, we propose a new similarity metric that jointly considers the node attributes and topology structure, guiding the node aggregation process. Secondly, a similarity feature smoothing strategy based on graph Laplacian filters is employed to suppress the noise and reduce error accumulation caused by local inconsistencies and temporal fluctuations. Thirdly, a temporal contrastive learning module is designed to accurately capture the evolution of node representations. It first fuses the local and global structural features to overcome the limitations of single-perspective learning, and then incorporates a Long Short-Term Memory (LSTM)-based temporal dynamics modeling strategy to capture the evolutionary patterns of node representations. Finally, the learned representations are clustered using the K-means algorithm to achieve accurate community detection. Extensive experiments on six benchmark adaptive oscillatory networks demonstrate the effectiveness and robustness of the proposed SSGCL method.