Interpretable Causal Discovery Using Causal-Effect Constraints

Causal discovery aims to identify the underlying causal relationships within a system based on observed data. Beyond simply predicting causal edges, there's a growing need to interpret and explain observed or hypothesized phenomena, such as particularly strong causal effects. This research addresses the task of conditional causal discovery, framing it as a Bayesian inference problem where the goal is to target the posterior over causal graphs and parameters, conditional on specific events like a causal-effect constraint. A significant computational challenge arises when the conditioning event has a small posterior mass, making existing Bayesian causal discovery approaches struggle. To overcome this, the proposed method adapts rare-event estimation techniques. It performs inference in the joint graph-parameter space by gradually guiding a particle population towards the constrained region while simultaneously maintaining samples that accurately approximate the conditional posterior. Empirical evaluations on synthetic graphs confirmed the accuracy of this approach across various scales. A case study using the Sachs protein dataset further demonstrated how the method can facilitate scientific exploration by providing pathway-level summaries, offering a powerful tool for understanding complex causal mechanisms.