<ul><li>Determining conditional independence (CI) relationships between random variables is a fundamental task in machine learning and statistics, particularly in high-dimensional settings.</li><li>Existing generative model-based CI testing methods, like those using generative adversarial networks (GANs), face challenges with modeling conditional distributions and training instability, leading to lower performance.</li><li>A novel CI testing method utilizing score-based generative modeling is proposed, ensuring precise Type I error control and strong testing power.</li><li>The method involves a sliced conditional score matching scheme for accurate estimation, Langevin dynamics conditional sampling for sample generation, and a goodness-of-fit stage for verification, outperforming current methods in experiments.</li></ul>

Score-based Generative Modeling for Conditional Independence Testing

Discover more