Algorithmic Recourse in Partially and Fully Confounded Settings Through Bounding Counterfactual Effects

Algorithmic recourse aims to provide actionable recommendations to\nindividuals to obtain a more favourable outcome from an automated\ndecision-making system. As it involves reasoning about interventions performed\nin the physical world, recourse is fundamentally a causal problem. Existing\nmethods compute the effect of recourse actions using a causal model learnt from\ndata under the assumption of no hidden confounding and modelling assumptions\nsuch as additive noise. Building on the seminal work of Balke and Pearl (1994),\nwe propose an alternative approach for discrete random variables which relaxes\nthese assumptions and allows for unobserved confounding and arbitrary\nstructural equations. The proposed approach only requires specification of the\ncausal graph and confounding structure and bounds the expected counterfactual\neffect of recourse actions. If the lower bound is above a certain threshold,\ni.e., on the other side of the decision boundary, recourse is guaranteed in\nexpectation.\n