Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The authors identify that decoupled diffusion forecasters suffer from a deterministic-to-residual structural alignment problem, wherein dependency graphs extracted from deterministic predictions misrepresent the true dependency structure of stochastic residuals. To resolve this, they formulate GARDiff, a framework that dynamically adapts deterministic graph priors during residual diffusion using residual uncertainty estimation and reverse-process edge sparsification. Across six real-world multivariate time-series benchmarks, this adaptive alignment yields consistent gains in probabilistic forecast accuracy and uncertainty calibration over state-of-the-art diffusion baselines.
Deterministic dependency graphs severely misalign with stochastic residual structures in decoupled diffusion forecasters, but dynamically sparsifying edges via residual uncertainty estimation eliminates redundant conditioning and repairs calibration.
Diffusion models have recently shown strong potential for probabilistic multivariate time-series forecasting by modeling complex conditional distributions. Recent decoupled diffusion frameworks further separate forecasting into deterministic prediction and stochastic residual generation, making it natural to derive dependency graphs from deterministic representations and use them to guide residual diffusion. However, we show that this direct structural transfer is unreliable. Although deterministic-derived graphs encode useful global dependency priors, they exhibit substantial edge-level misalignment with residual dependency structures, introducing inaccurate or redundant conditions during residual generation. This reveals a previously overlooked deterministic-to-residual structural alignment problem in decoupled diffusion forecasting. To address this problem, we propose GARDiff, a Graph-Aligned Residual Diffusion framework for probabilistic multivariate time-series forecasting. Instead of treating deterministic-derived graphs as fixed diffusion conditions, GARDiff progressively adapts them to residual generation. Specifically, GARDiff estimates residual uncertainty to distinguish high- and low-uncertainty regions, enabling uncertainty-aware structural refinement, and further performs timestep-aware edge sparsification during reverse diffusion to evolve graph conditions from broad dependency aggregation to localized residual refinement. Extensive experiments on six real-world benchmarks demonstrate that GARDiff consistently improves probabilistic forecasting performance and uncertainty calibration over strong baselines.