Skip to main content
Back to timeline
Communications in computer and information scienceSource publication:

Causal recourse stops assuming the causal model is known: ten user queries bring cost close to ground truth on linear models

Synopsis

When a machine-learning model makes an unfavorable decision, algorithmic recourse must tell the person what to change, yet existing causal recourse methods assume the causal model governing individual features is already known. A human-in-the-loop framework instead asks the user a series of soft-intervention questions and uses Bayesian inference on the noisy answers to estimate that user's structural causal model, then derives the cheapest recourse action from it. On linear synthetic models, recourse from the estimated model tracks the ground-truth model in validity and cost and beats a prior-only non-causal baseline; on a non-linear model validity drops while cost stays below the baseline.

AI-generated editorial illustration: Personalized Causal Recourse: A Human-In-The-Loop Approach

Interpretation

Recourse recommendations are generated around each user's own causal structure instead of a shared, assumed global causal model. Earlier causal recourse methods assumed the true causal model of an individual's features, both the causal graph and the structural equations, was known or reliably approximable from observational data, although causal mechanisms may differ between individuals and observational data may be unavailable. Across three three-variable synthetic causal models, actions found under the estimated model were applied to the ground-truth model: for the linear model, validity and cost of estimated-model recourse closely tracked the ground truth, while a non-causal baseline built only from prior constraints was consistently worse.

The user is treated as an imperfect source of causal knowledge, and the system approaches that person's causal model by asking what happens after interventions. Each round draws a random soft intervention, adding an offset to one variable, and asks the user to report the resulting values of descendant variables, forming an intervention-response dataset. Structural coefficients are inferred by Markov chain Monte Carlo posterior sampling with ten interventions per run, and user answers are modeled as a mixture that lets a fraction come from perturbed, incorrect causal parameters; on synthetic data, linear-model estimation error fell as interventions increased, with most gains within the first five to six.

Even an imperfect estimated model concentrates recourse actions on variables close to those the ground-truth model would choose. A prior-only baseline spread its intervention targets more diffusely and less optimally. On the two linear synthetic models, the estimated model's intervention-target distribution sat closer to the ground truth than the prior baseline, especially in low-uncertainty settings; intervention patterns for the non-linear model were not analyzed because its validity results were fragile.

Perspective

The demonstrated scope is three three-variable synthetic structural causal models with simulated rather than real user responses, evaluated by applying actions from the estimated model to the ground-truth model and measuring recourse validity and cost. The authors present the framework as a proof of concept within this constrained model class and note that scaling to higher-dimensional or more complex settings will be challenging.

User responses are modeled as a mixture that allows a fraction to come from perturbed, incorrect causal parameters, but the authors state this does not capture structured biases, anchoring effects, or internally coherent yet systematically wrong causal beliefs. Estimation accuracy is sensitive to user error rate, tolerance threshold and robustness requirement, and greater robustness uniformly lowers validity; even with the ground-truth model, recourse is not guaranteed for every negatively classified individual.