KoopmanGraph¶
KoopmanGraph integrates Graph Neural Networks with Koopman operator theory for spatiotemporal graph dynamics. The library lifts node features into a latent space with topology-aware encoders, advances state via a learned finite-dimensional Koopman operator, and decodes predictions back to physical node features.
Version 0.2 adds spectral analysis, model checkpointing, temporal evaluation metrics, operator stability options, edge-weight support, classical DMD/EDMD/DMDc baselines, Koopman-with-control dynamics, dynamic-topology sequences, and advanced training utilities including windowed mini-batching.
Contents