Causal Time Series Generation via Diffusion Models

2025-09-25

AI Summary

Existing time-series generation (TSG) models capture only observational correlations among covariates, neglecting unobserved confounders, and thus lack support for causal intervention and counterfactual reasoning. This work pioneers the integration of causal inference into TSG, introducing the novel task family of Causal Time-Series Generation (CaTSG), encompassing both interventional and individualized counterfactual scenarios. We propose a unified diffusion-based framework featuring a causal score function and an abduction-action-prediction mechanism, enabling causal-guided sampling via backdoor adjustment. Our approach transcends traditional correlational modeling by explicitly accounting for latent confounding. Empirically, it achieves significant improvements in generation fidelity on both synthetic and real-world benchmarks. Moreover, it is the first method to enable interpretable intervention-response analysis and individual-level counterfactual sequence generation.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• 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
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Time series generation (TSG) synthesizes realistic sequences and has achieved remarkable success. Among TSG, conditional models generate sequences given observed covariates, however, such models learn observational correlations without considering unobserved confounding. In this work, we propose a causal perspective on conditional TSG and introduce causal time series generation as a new TSG task family, formalized within Pearl's causal ladder, extending beyond observational generation to include interventional and counterfactual settings. To instantiate these tasks, we develop CaTSG, a unified diffusion-based framework with backdoor-adjusted guidance that causally steers sampling toward desired interventions and individual counterfactuals while preserving observational fidelity. Specifically, our method derives causal score functions via backdoor adjustment and the abduction-action-prediction procedure, thus enabling principled support for all three levels of TSG. Extensive experiments on both synthetic and real-world datasets show that CaTSG achieves superior fidelity and also supporting interventional and counterfactual generation that existing baselines cannot handle. Overall, we propose the causal TSG family and instantiate it with CaTSG, providing an initial proof-of-concept and opening a promising direction toward more reliable simulation under interventions and counterfactual generation.

Problem

• Generating time series with causal relationships beyond observational correlations
• Extending time series generation to interventional and counterfactual settings
• Developing diffusion models that causally steer sampling using backdoor adjustment

Innovation

• Causal diffusion framework with backdoor-adjusted guidance
• Derives causal score functions via backdoor adjustment
• Supports observational, interventional, and counterfactual generation