UAI 2026 paper uses outlier events to falsify causal graphs

Determining true causal relationships from data remains a fundamental challenge in machine learning and statistics. A new paper accepted at the 42nd Conference on Uncertainty in Artificial Intelligence (UAI 2026) tackles this by turning outlier events into a diagnostic tool. Authors William Roy Orchard, Philipp M. Faller, and Dominik Janzing propose a method to falsify candidate causal graphs based on whether they can explain the propagation of an outlier event. Their key insight is a simple principle: weak outliers rarely cause strong outliers. By reversing the typical root cause analysis approach, they use observed outlier patterns to test if a proposed causal graph is consistent with the data.

The team presents what they call the first statistical tests specifically designed to assess the hypothesis that a candidate graph is the true causal graph. These tests offer false positive control and power guarantees against incorrect graphs—meaning they can reliably reject wrong models without incorrectly discarding the correct one. Notably, the method can operate with just a single outlier sample, making it practical for scenarios with limited data. This work has immediate implications for fields where ground truth causal structures are unknown, such as epidemiology, economics, and systems biology, providing a rigorous way to validate or discard causal models before using them for decision-making.