Neural Computing and Applications Программное обеспечение
DSE-SBM: dynamic structure evolution-based spatio-temporal graph learning for IoT state modeling and decision support
Аннотация
Large-scale IoT systems often operate under dynamic environments with communication variability and evolving device interactions, making reliable system modeling challenging. Modern IoT-enabled networks generate complex spatio-temporal data with dynamically evolving interdependencies, making accurate state modeling and decision support challenging. Existing methods typically rely on static network structures or predefined dependencies, which fail to capture dynamic interactions among distributed devices and system nodes. Additionally, they often overlook behavior consistency in operational states, leading to predictions that may violate physical or functional constraints. To address these issues, this paper proposes Dynamic Structure Evolution for System Behavior Modeling (DSE-SBM). The framework models inter-node dependencies as adaptive, learnable structures and introduces a dynamic graph evolution mechanism to reconstruct network topology in response to changing system states and operational conditions. A behavior consistency regularization strategy is incorporated to enforce structural and temporal coherence, ensuring reliable performance. To support large-scale deployment, a structure-aware distributed learning scheme enables collaborative modeling across decentralized nodes while preserving heterogeneity. Experimental results on six real-world datasets show that DSE-SBM consistently achieves the best overall performance, for example reducing RMSE from 6.67 to 6.21 on METR-LA and from 5.12 to 4.76 on PEMS04 compared with the strongest baseline, while also improving stability and robustness under heterogeneous settings.