We study the problem of discovering reliable causal rules from observational data. Traditional descriptive rule discovery techniques do not suffice to this end, as they struggle with the consistent detection of (potentially rare) conditions that have a strong effect on an output variable of interest. Among the sources of inconsistency are that naive empirical effect estimations have a high variance, and, hence, their maximization is highly optimistically biased unless the search is artificially restricted to high frequency events. Secondly, observational effect measurements are often highly unrepresentative of the underlying causal effect because they are skewed by the presence of confounding factors. This is a concern especially in scientific data analysis. To address these issues, we present a novel descriptive rule discovery approach based on reliably estimating the conditional effect given the potential confounders. We demonstrate that the corresponding score is a conservative and consistent effect estimator, identify the admissible data generation process under which causal rule discovery is possible, and derive an efficient optimization algorithm that successfully detects valuable rules on a multitude of real datasets. Important for both causal and associational data exploration, the presented approach naturally allows for iterative rule discovery, where new non-redundant rules can be found by treating previously discovered rules as confounders in subsequent iterations.
NeurIPS 2018 workshop on Causal Learning