DiffCATS: Causally Associated Time-Series Generation Through Diffusion Models

Abstract

Modeling and recovering causal relationships in time-series data can be crucial for supporting real-world interventions and decision-making, but progress in Time-Series Causal Discovery (TSCD) is often limited by the lack of high-quality datasets with diverse and realistic temporal causal relationships. This highlights the need to provide synthetic time-series generation tools, with realism as a primary objective, an aspect that requires incorporating causal relationships beyond mere correlation. To address this challenge, we propose a diffusion model called DiffCATS. It simultaneously generates multiple causally associated time-series as well as a ground truth causal graph that reflects their mutual temporal dependencies, requiring only observational time-series data for training. Experiments demonstrate that it outperforms state-of-the-art methods in producing realistic time-series with causal graphs that closely resemble those of real-world phenomena. We highlight the practical utility of our data on three downstream tasks, including benchmarking widely used TSCD algorithms.