DoFlow: Causal Generative Flows for Interventional and Counterfactual Time-Series Prediction

2025-11-04

AI Summary

This work addresses causal forecasting in multivariate time series. We propose the first generative modeling framework that jointly integrates causal directed acyclic graphs (DAGs) with continuous normalizing flows (CNFs). Built upon an encoder-decoder architecture, our method unifies observational forecasting, interventional inference, and counterfactual querying, while providing explicit likelihood estimation of future trajectories—enabling both system-level forecasting and likelihood-based anomaly detection. We theoretically establish the feasibility of counterfactual recovery. Empirical evaluation on synthetic data and real-world time series—including hydropower station operations and cancer treatment records—demonstrates substantial improvements in multi-step forecasting accuracy, faithful responses to diverse causal queries (observational, interventional, counterfactual), and effective identification of anomalous trajectories. Our approach delivers an interpretable, generalizable, and probabilistically principled solution for causal inference in time series.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Action, Change, and Causality

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: Attacks and countermeasures in recommendation systems

Abstract

Time-series forecasting increasingly demands not only accurate observational predictions but also causal forecasting under interventional and counterfactual queries in multivariate systems. We present DoFlow, a flow based generative model defined over a causal DAG that delivers coherent observational and interventional predictions, as well as counterfactuals through the natural encoding and decoding mechanism of continuous normalizing flows (CNFs). We also provide a supporting counterfactual recovery result under certain assumptions. Beyond forecasting, DoFlow provides explicit likelihoods of future trajectories, enabling principled anomaly detection. Experiments on synthetic datasets with various causal DAG and real world hydropower and cancer treatment time series show that DoFlow achieves accurate system-wide observational forecasting, enables causal forecasting over interventional and counterfactual queries, and effectively detects anomalies. This work contributes to the broader goal of unifying causal reasoning and generative modeling for complex dynamical systems.

Problem

• Predicts time-series under interventions and counterfactual scenarios
• Unifies causal reasoning with generative flow-based modeling
• Provides explicit likelihoods for anomaly detection in forecasts

Innovation

• Flow-based generative model over causal DAG
• Continuous normalizing flows for counterfactual predictions
• Explicit likelihood computation for anomaly detection