algorithmic recourse
In plain wordsGiving a person who received an unfavorable automated decision a realistic set of actions they could take to get a better result next time, such as paying down a particular debt.
The prospect that someone who receives an unfavorable automated decision can turn it around by steps actually open to them, and the methods that find such steps. Recourse methods search for low-cost changes restricted to features the person can actually alter (savings, not age or ethnicity), respecting bounds and the differing costs of actions; causal formulations model actions as interventions, since changing one feature (education) also changes others (income). Recourse extends counterfactual explanationscounterfactual explanation, n.: An explanation that tells someone the smallest change that would have produced a different outcome — for example, that the application would have been accepted if it had listed two more years of work experience. from describing the model to advising the person, and raises its own fairness question: if one group must make costlier changes than another to obtain the same outcome, the system offers unequal recourse even when its error rates are equal. Recourse can also be invalidated when the model is retrained or when people strategically adapt.
The audit found that applicants over 60 faced far costlier algorithmic recourse than younger applicants.
Origin
Introduced as a machine-learning problem by Berk Ustun, Alexander Spangher and Yang Liu in 'Actionable Recourse in Linear Classification' (ACM FAT* 2019); Amir-Hossein Karimi, Bernhard Schölkopf and Isabel Valera reframed it causally in 'Algorithmic Recourse: from Counterfactual Explanations to Interventions' (FAccT 2021).
Usage note
Recourse can conflict with other goals: disclosing how to flip a decision invites gaming, and guaranteeing recourse constrains which models can be deployed.
Related
- counterfactual explanationcounterfactual explanation, n.: An explanation that tells someone the smallest change that would have produced a different outcome — for example, that the application would have been accepted if it had listed two more years of work experience.
- algorithmic fairnessalgorithmic fairness, n.: The effort to make automated decisions fair: to say in numbers what fair treatment means, measure whether a system achieves it, and correct the system if it does not.
- right to explanationright to explanation, n.: The idea that a person affected by a decision made or shaped by an algorithm — a refused loan, a rejected job application — is entitled to a meaningful account of why.
- local explanationlocal explanation, n.: An explanation of why an AI model gave one particular answer for one particular case — this patient, this loan — rather than how it behaves in general.
- protected attributeprotected attribute, n.: A personal trait, such as race, sex, religion, age or disability, that the law says must not be used to treat people unfairly.
Sources
- Ustun, B., Spangher, A. & Liu, Y. (2019). Actionable Recourse in Linear Classification. Proceedings of ACM FAT* 2019.
- Karimi, A.-H., Schölkopf, B. & Valera, I. (2021). Algorithmic Recourse: from Counterfactual Explanations to Interventions. Proceedings of ACM FAccT 2021.
Mentioned in
counterfactual explanationcounterfactual explanation, n.: An explanation that tells someone the smallest change that would have produced a different outcome — for example, that the application would have been accepted if it had listed two more years of work experience.