testing causal graphs with outlier events

source: arxiv statistics ml: falsifying causal graphs with outlier events

level: research

true causal relationships are rarely known, and inferring causal graphs from data is hard. a key challenge is evaluating a candidate causal graph without ground truth. this work proposes falsifying candidate graphs by checking if they can explain how an outlier event spreads. the core idea is that weak outliers rarely cause strong ones. if a graph implies an outlier propagation pattern that does not match the data, the graph is likely wrong.

the approach turns a known root cause analysis principle on its head. instead of using the weak-outlier principle to find root causes without a graph, it uses the principle to test a given graph. the method introduces statistical tests for the hypothesis that a candidate graph is the true causal graph. these tests control false positives, guarantee power against incorrect graphs, and can work with just a single outlier sample.

this provides a practical way to reject bad causal models using observed outlier events. it does not require knowing the true graph beforehand. the tests are designed to be reliable even with limited data, making them useful in real-world settings where outliers occur naturally. the work opens a new direction for causal validation by focusing on falsification rather than full discovery.

why it matters: it gives data scientists a simple, statistically sound way to check if a causal model is wrong using outlier events, which are common in real data.

source: arxiv statistics ml: falsifying causal graphs with outlier events