Semantically-Guided Inference for Conditional Diffusion Models: Enhancing Covariate Consistency in Time Series Forecasting

TL;DR

This paper introduces SemGuide, a versatile inference-time technique that improves covariate consistency in conditional diffusion models for time series forecasting, leading to better alignment and accuracy.

Contribution

SemGuide is a novel, model-agnostic inference method that enhances semantic alignment in conditional diffusion models without retraining, using a scoring network and importance reweighting.

Findings

Improves covariate alignment in time series forecasts.

Enhances predictive accuracy under complex conditions.

Compatible with any conditional diffusion model.

Abstract

Diffusion models have demonstrated strong performance in time series forecasting, yet often suffer from semantic misalignment between generated trajectories and conditioning covariates, especially under complex or multimodal conditions. To address this issue, we propose SemGuide, a plug-and-play, inference-time method that enhances covariate consistency in conditional diffusion models. Our approach introduces a scoring network to assess the semantic alignment between intermediate diffusion states and future covariates. These scores serve as proxy likelihoods in a stepwise importance reweighting procedure, which progressively adjusts the sampling path without altering the original training process. The method is model-agnostic and compatible with any conditional diffusion framework. Experiments on real-world forecasting tasks show consistent gains in both predictive accuracy and covariate alignment, with especially strong performance under complex conditioning scenarios.