Dynamic Community Detection using Multi-View Graph Representation and Temporal Contrastive Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Dynamic community detection is of great importance to understand the evolution of relationships among entities in social networks over time. However, existing methods rely on single-view graph information, ignoring the global structural and temporal information. To address this issue, this paper presents a dynamic community detection framework which contains two main stages. The first stage includes processing the social network data by using the Improved Graph Autoencoder (IGAE), which optimizes the standard GAE using the Sea Lion Optimization Algorithm (SLOA) to get informative representations for nodes. The second stage of the method is called the Dynamic Multi-View Temporal Graph Contrastive Network (DMVT-GCN) and is responsible for processing of dynamic graphs. Firstly, the framework uses neighbour overlap similarity and topological structure similarity to improve the graph feature representation. Then, the Improved Graph Isomorphism Networks (I-GIN) is introduced to learn both local node representations and global graph information. The jumping knowledge network (JK-Net) is integrated into the GIN model to perform adaptive aggregation. A Dilated Coupled-Gate Temporal Memory Network (DCGTMNet)-based temporal dynamics modelling module is incorporated to update the parameters of I-GINin each time step. The DCGTMNet integrates Coupled Input–Forget Gate Long Short-Term Memory (CIFG-LSTM)with structured dilated temporal connections to enable long-range dependency learning. Additionally, a Dual-Scale Graph Contrastive Learning (DS-GCL) based network smoothing strategy is designed to improve the consistency and robustness of the node representations. The obtained node representations are input into the Embedding-Guided Adaptive Louvain Refinement (EGALR) algorithm for community detection at different time steps. Extensive experimental results demonstrate that the proposed framework achieves superior performance, achieving Normalized Mutual Information (NMI) of 0.9228, 0.8750, 0.8849, and 0.9049 on the Cora, HighSchool, Facebook, Bitcoin-OTC datasets, respectively.