Preprint
Aug 2026
Geometry-aware neural causal discovery for large-scale spatiotemporal systems
GeoDCD is introduced, a geometry-aware neural framework that uses spatial coordinates to initialize a learnable hierarchy and converts a trained nonlinear predictor into time-varying directed graphs through input-output Jacobian sensitivity analysis.
Hao-Yang Yan, Kai-Qi Zhao, Weiping Wang et al.
· 0 citations