STR-GNN: stability-regularized graph neural networks for suppressing spurious temporal variations in dynamic community detection

Abstract

Abstract The study of temporal graphs frequently encounters spurious temporal fluctuations, wherein transient noise, inadequate observations, or ephemeral structural perturbations result in unstable and inconsistent community assignments. In dynamic networks, such variations may stem from measurement inaccuracies, sampling biases, or sudden yet non-informative changes in topology, leading graph neural networks (GNNs) to excessively respond to local fluctuations rather than accurately reflecting the genuine growth of communities. Consequently, the acquired representations may demonstrate considerable temporal inconsistency, resulting in community structures that fluctuate erratically across successive time intervals. This instability significantly constrains the reliability of dynamic community recognition in temporal GNNs, especially in streaming, partially observed, or weakly supervised contexts, where the model must perpetually adjust to changing graph structures without comprehensive ground-truth supervision. In these circumstances, the learning process may provide community assignments that seem credible at each moment yet lack overall temporal consistency. We refer to this issue as spurious temporal variation, where temporal GNNs produce unstable community assignments caused by transient perturbations rather than genuine structural evolution. The suggested method mitigates false temporal fluctuations by implementing stability-aware regularization and consistency restrictions across successive graph snapshots, promoting the model’s ability to maintain coherent community structures while adapting to authentic structural changes. The proposed framework enhances the resilience, reliability, and interpretability of temporal GNN-based community identification by stabilizing community evolution over time, hence rendering it more appropriate for real-world dynamic networks characterized by noise, perturbations, and incomplete data.