Conceptos Básicos
DiffSTG combines STGNN capabilities with DDPM uncertainty measurements for improved probabilistic forecasting.
Resumen
Introduction: Discusses the need for probabilistic STG forecasting.
Background: Explains limitations of deterministic models and diffusion-based probabilistic models.
DiffSTG Formulation: Introduces conditional diffusion model and UGnet denoising network.
DiffSTG Implementation: Details UGnet architecture and sampling acceleration strategy.
Experiments: Compares DiffSTG with baselines on real-world datasets, showing superior performance.
Inference Time: Reports inference speed of DiffSTG compared to TimeGrad and CSDI.
Visualization: Provides visual comparisons of predicted distributions by DiffSTG and other methods.
Estadísticas
DiffSTGはCRPSを5.6%、4.3%、および14.3%削減しました。