BM-AN: Uncertainty-Aware Dynamic Community Detection via Bayesian Markov Attention Networks with Graph Memory

Dynamic community discovery in temporal networks is fundamentally hindered by noise amplification, structural drift, andthe absence of dependable uncertainty estimation. Current methodologies generally depend on deterministic temporalsmoothing or snapshot-based optimization, resulting in unstable community trajectories and overly confident forecasts innon-stationary dynamics. We present Bayesian Markov Attention Networks with Graph Memory (BM-AN), a probabilisticframework that conceptualizes dynamic community discovery as latent-variable inference on changing graphs. BM-AN employsBayesian parameter inference to assess epistemic uncertainty in community assignments, a hybrid attention mechanismto simultaneously capture global context and local topology, and a memory-augmented Markov regularization scheme thatensures both short-term temporal coherence and long-term structural persistence. In contrast to previous temporal GNNsthat implicitly smooth representations, BM-AN actively regulates community evolution at both distributional and spectral levels,mitigating erroneous high-frequency oscillations while maintaining significant structural transitions. This approach producesstable and interpretable community trajectories, even in the presence of significant noise, disturbance, and resource limitations.Comprehensive investigations on various synthetic and real-world dynamic networks reveal that BM-AN routinely surpassesleading baselines in modularity, temporal stability, evolutionary consistency, and uncertainty calibration. The findings confirmthat uncertainty-aware temporal inference is essential for dependable dynamic community detection.