Quantifying Consistency and Information Loss for Causal Abstraction Learning


In this presentation we quickly review the idea of defining an abstraction between structural causal models and we present the standard measure of abstraction error proposed in the literature. We then consider some potential limitations when using this single measure to assess or learn abstractions. To overcome this limit, we propose an extension of the original definition of abstraction approximation, we derive new measures of abstraction error, and we discuss theoretical and applied properties of these new measures.

Slides here