Abstract
To effectively study complex causal systems, it is often useful to constructabstractions of parts of the system by discarding irrelevant details whilepreserving key features. The Information Bottleneck (IB) method is a widelyused approach to construct variable abstractions by compressing randomvariables while retaining predictive power over a target variable. Traditionalmethods like IB are purely statistical and ignore underlying causal structures,making them ill-suited for causal tasks. We propose the Causal InformationBottleneck (CIB), a causal extension of the IB, which compresses a set ofchosen variables while maintaining causal control over a target variable. Thismethod produces abstractions of (sets of) variables which are causallyinterpretable, give us insight about the interactions between the abstractedvariables and the target variable, and can be used when reasoning aboutinterventions. We present experimental results demonstrating that the learnedabstractions accurately capture causal relations as intended.