Predicting counterfactual distributions is crucial for scientific modeling and decision-making in fields like public health and medicine, where understanding alternative scenarios can inform better interventions. Traditional methods often depend on point estimates or purely data-driven models, which struggle when data is limited. To address this, the authors propose ODE-Diff, a diffusion-based time series framework that leverages imperfect expert models by extracting high-level signals to act as structured priors in generative modeling. This hybrid approach effectively combines mechanistic insights with data-driven techniques, enhancing the interpretability and reliability of causal inferences. The framework is evaluated on semi-synthetic COVID-19 simulations, synthetic pharmacological dynamics, and real-world case studies, consistently outperforming strong baseline models in both point prediction accuracy and distributional fidelity. The integration of expert knowledge helps mitigate risks associated with data scarcity and model misspecification, making ODE-Diff a robust tool for counterfactual analysis. This method holds promise for improving decision-making in critical domains by providing more nuanced and trustworthy predictions. Future work could explore further refinement of expert model integration and broader application across diverse dynamical systems.
👉 Pročitaj original: arXiv AI Papers
