| Conditional independence testing is fundamental in predictive and causal modeling. While numerous tests have been proposed in the literature, many suffer from limitations such as strong parametric assumptions, restrictions on variable types, or prohibitive computational cost—making them ill-suited for application to complex, real-world datasets. Recent non-parametric approaches address these issues by testing conditional independence based on predictive performance. The core idea is that if two variables X and Y are conditionally dependent given a set Z, then a predictive model using both X and Z to predict Y should outperform a model that uses only Z. Building on this principle, we propose a non-parametric conditional independence test that leverages the capabilities of Automated Machine Learning (AutoML). We design an architecture that not only selects and tunes predictive models, but also ensures the statistical robustness of the test. Unlike previous methods, our approach maintains proper test calibration (Type I error control) and power, and minimizes computational cost. In addition, it is robust to different data types (including mixed data), functional relationships, and large sample sizes or conditioning sets. Through extensive experiments on both synthetic and real-world datasets, we demonstrate the practical applicability of our method and its competitive performance relative to state-of-the-art alternatives. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.