Reproduce “Exactly Computing do-Shapley Values” on OpenResearch

About this paper

R. Teal Witter, Álvaro Parafita, Tomas Garriga, Maximilian Muschalik, Fabian Fumagalli, Axel Brando, Lucas Rosenblatt

Researchers from Claremont McKenna College, Barcelona Supercomputing Center, Novartis, LMU Munich, and NYU developed methods to exactly compute or more accurately estimate do-Shapley values for Structural Causal Models. Their approach reformulates do-Shapley values over "irreducible sets" of interventions and introduces an efficient identifiability check, leading to significantly reduced computational costs and improved estimation accuracy, especially in resource-constrained environments.

Key insights

• Many causal interventions in Structural Causal Models (SCMs) are functionally redundant, allowing do-Shapley values to be expressed more compactly as a sum over a potentially much smaller number of unique causal effects or "irreducible sets" (r) than 2^d.
• By leveraging the underlying graph structure of SCMs through boundary sampling, estimators can efficiently discover distinct causal effects, leading to substantially higher accuracy and seamless convergence to exact values as the computational budget increases.
• The non-parametric identifiability of all possible do-interventions relevant to a do-Shapley value can be globally verified by only checking the identifiability of each of the *d* individual singleton interventions.

Summary by alphaXiv.