Reproduce “Causal Entropy and Information Gain for Measuring Causal Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Francisco Nunes Ferreira Quialheiro Simoes, Mehdi Dastani, Thijs van Ommen
Researchers from Utrecht University introduce 'causal entropy' and 'causal information gain,' novel information-theoretic measures quantifying how much control intervening on a feature provides over an outcome variable. These measures distinguish genuine causal influence from statistical correlations, addressing a limitation in current interpretable AI methods.
Key insights
- Causal information gain (I_c) provides a robust, information-theoretic metric for quantifying the degree of control a feature provides over an outcome variable, rather than just statistical association.
- A non-zero causal information gain (I_c ≠ 0) definitively implies the existence of a total causal effect, offering a clear criterion for identifying genuine causal influence.
- Unlike standard mutual information, causal information gain is inherently asymmetric (at least one of I_c(Y | do(X~X')) or I_c(X | do(Y~Y')) must be zero), reflecting the directional nature of causality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.