[2309.07703] Causal Entropy and Info Acquire for Measuring Causal Management

Obtain a PDF of the paper titled Causal Entropy and Info Acquire for Measuring Causal Management, by Francisco Nunes Ferreira Quialheiro Simoes and a couple of different authors

Obtain PDF

Summary:Synthetic intelligence fashions and strategies generally lack causal interpretability. Regardless of the developments in interpretable machine studying (IML) strategies, they often assign significance to options which lack causal affect on the end result variable. Deciding on causally related options amongst these recognized as related by these strategies, and even earlier than mannequin coaching, would supply an answer. Characteristic choice strategies using data theoretical portions have been profitable in figuring out statistically related options. Nonetheless, the knowledge theoretical portions they’re based mostly on don’t incorporate causality, rendering them unsuitable for such situations. To handle this problem, this text proposes data theoretical portions that incorporate the causal construction of the system, which can be utilized to judge causal significance of options for some given consequence variable. Particularly, we introduce causal variations of entropy and mutual data, termed causal entropy and causal data acquire, that are designed to evaluate how a lot management a function gives over the end result variable. These newly outlined portions seize modifications within the entropy of a variable ensuing from interventions on different variables. Elementary outcomes connecting these portions to the existence of causal results are derived. The usage of causal data acquire in function choice is demonstrated, highlighting its superiority over commonplace mutual data in revealing which options present management over a selected consequence variable. Our investigation paves the way in which for the event of strategies with improved interpretability in domains involving causation.