ChaosNetBench: Benchmarking Spatio-Temporal Graph Neural Networks on Chaotic Lattice Dynamics
Researchers have introduced ChaosNetBench (CNB), a novel synthetic benchmark dataset and evaluation framework designed to assess the performance of Spatio-temporal Graph Neural Networks (STGNNs) under controlled chaotic conditions. Addressing limitations in current real-world benchmarks, CNB utilizes a lattice of coupled standard maps with tunable parameters for local chaos, coupling strength, and system size. The framework includes 96 system instances and 9,600 trajectories, offering known topology and dynamics for rigorous analysis. The study evaluates 13 architectures, including five STGNNs and eight non-graph baselines like TCN and N-BEATS. Results indicate a regime-dependent transition: non-graph models remain competitive in low-chaos environments, while STGNNs such as Graph WaveNet and D2STGNN demonstrate superior resilience in high local and global chaos scenarios. This work provides a reusable testbed for systematically comparing architectural capacities in handling complex dynamical systems, aiming to improve forecasting accuracy in fields like traffic and weather modeling.
Wire timeline
ChaosNetBench: Benchmarking Spatio-Temporal Graph Neural Networks on Chaotic Lattice Dynamics
Researchers have introduced ChaosNetBench (CNB), a novel synthetic benchmark dataset and evaluation framework designed to assess the performance of Spatio-temporal Graph Neural Networks (STGNNs) under controlled chaotic conditions. Addressing limitations in current real-world benchmarks, CNB utilizes a lattice of coupled standard maps with tunable parameters for local chaos, coupling strength, and system size. The framework includes 96 system instances and 9,600 trajectories, offering known topology and dynamics for rigorous analysis. The study evaluates 13 architectures, including five STGNNs and eight non-graph baselines like TCN and N-BEATS. Results indicate a regime-dependent transition: non-graph models remain competitive in low-chaos environments, while STGNNs such as Graph WaveNet and D2STGNN demonstrate superior resilience in high local and global chaos scenarios. This work provides a reusable testbed for systematically comparing architectural capacities in handling complex dynamical systems, aiming to improve forecasting accuracy in fields like traffic and weather modeling.
cs.AI updates on arXiv.org