GTENN: A Spatiotemporal Graph Neural Framework for Community Discovery in Dynamic Social Networks

TL;DR

GTENN tackles dynamic community discovery by integrating spatial graph structure with temporal evolution through a spatiotemporal embedding pipeline that combines GCNs, GRUs, and SOM clustering. It learns per-snapshot node embeddings with shared GCN weights across time and optimizes a distance-preserving, unsupervised objective while clustering with a Self-Organizing Map. Across synthetic LFR benchmarks and real-world networks, GTENN consistently outperforms static, dynamic, and attention-based baselines in purity and NMI, demonstrating robust detection of evolving communities with improved efficiency. The work highlights a scalable, snapshot-based framework for uncovering dynamic community structures and points to future extensions into continuous-time modeling and broader application domains.

Abstract

Community discovery is one of the key issues in the study of dynamic social networks. Traditional community discovery algorithms mainly focus on the formation and dissolution of links between nodes, and thus fail to capture richer spatial and temporal patterns underlying network evolution. To address this limitation, we propose GTENN, a spatiotemporal graph neural framework for community discovery in dynamic social networks. GTENN integrates spatial structure and temporal dynamics within a unified embedding architecture. First, Graph Convolutional Networks (GCN) are employed to aggregate latent spatial information and learn expressive node representations at each snapshot. Next, Gated Recurrent Units (GRU) are used to model temporal evolutions of node embeddings, effectively capturing node dynamism and relationship propagation across time. Finally, a Self-Organizing Map (SOM) is applied to the learned spatiotemporal embeddings to cluster nodes and infer their community affiliations. We conduct experiments on four types of dynamic networks, and the results show that GTENN consistently outperforms traditional community discovery algorithms in terms of purity, normalized mutual information, homogeneity, and completeness. These findings demonstrate the superior ability of GTENN to accurately uncover evolving community structures hidden in dynamic social networks.

Figures (7)

• Figure 1: Framework of the proposed GTENN algorithm. The process integrates three main components: (i) spatial feature extraction using GCNs to learn node representations from network topology, (ii) temporal feature extraction via GRU to capture node dynamics and long-term dependencies by sharing GCN weights across time, and (iii) community discovery using SOM to analyze the learned embeddings and identify evolving communities in dynamic networks.
• Figure 2: Spatial feature extraction stage of GTENN, where GCNs learn node features by aggregating structural information.
• Figure 3: Structure of the GRU used for temporal feature extraction in GTENN.
• Figure 4: Performance comparison of graph embedding methods on synthetic and real datasets. Results show that GTENN consistently achieves higher purity and NMI scores compared to static methods (SDNE, DeepWalk) and dynamic methods (DynAE, DynGEM).
• Figure 5: Comprehensive comparison of community discovery algorithms. GTENN demonstrates superior performance over traditional static (Louvain), dynamic (Dyperm, FacetNet), and embedding-based (DyCTWE) methods across both synthetic and real datasets.

Theorems & Definitions (4)

• Definition 1: Dynamic Network
• Definition 2: Dynamic Community
• Definition 3: Graph Embedding
• Definition 4: Adjacency Matrix