BTM-Net: Bayesian Temporal Memory Networks forUncertainty-Aware Dynamic Community Detection

Daozheng Qu, Yanfei Ma, Yibo Wang

Abstract: For the purpose of dynamic community detection in temporal graphs, it is necessary to simultaneously preserve stable community assignments in the presence of noisy and developing structures while also retaining sensitivity to true structural changes. Existing methods often either optimize each snapshot independently or enforce heuristic temporal smoothness. Furthermore, the majority of these methods do not explicitly describe the prediction uncertainty, which results in community trajectories that are either unstable or overconfident. We present an uncertainty-aware framework in which community assignments are described as probabilistic variables that evolve over time. This approach is intended to solve the constraints that have been identified. Bayesian inference is used to capture epistemic uncertainty, Markov temporal regularization is used to enforce short-term consistency, and a graph memory mechanism is used to retain long-range structural information. Additionally, the method combines all three of these techniques. It is possible to learn node representations that are capable of supporting strong community inference in contexts that are dynamic by utilizing a topology-aware encoder. Experiments conducted on synthetic and real-world temporal networks have shown that there is a consistent improvement in community quality, temporal stability, and predictive calibration when compared to representative baselines for static, dynamic, and temporal graph learning. According to the findings, it is essential to explicitly represent uncertainty in conjunction with multi-horizon temporal dependencies in order to achieve reliable and interpretable dynamic community detection.