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

2025-08-03

AI Summary

Diffusion models for time-series forecasting often suffer from semantic misalignment between generated trajectories and conditional covariates—especially in complex or multimodal settings—leading to degraded predictive performance. To address this, we propose a plug-and-play, inference-time semantic alignment method that requires no modification to the training procedure. Our approach introduces a lightweight semantic scoring network that dynamically evaluates stepwise alignment between sampled outputs and conditioning signals. Leveraging a proxy likelihood, it performs step-level importance reweighting to steer the diffusion sampling trajectory toward semantically consistent solutions. The method is model-agnostic and broadly compatible with existing diffusion-based forecasters. Extensive experiments on multiple real-world datasets demonstrate significant improvements in both forecasting accuracy and covariate alignment fidelity; gains are particularly pronounced under multimodal conditions, where semantic consistency is most challenging to maintain.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

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.

Problem

• Improves semantic alignment in diffusion-based time series forecasts
• Enhances covariate consistency without retraining diffusion models
• Addresses misalignment under complex multimodal conditioning scenarios

Innovation

• SemGuide enhances covariate consistency via scoring network
• Stepwise importance reweighting adjusts sampling path dynamically
• Model-agnostic plug-and-play method for diffusion models