Towards Realistic Individual Recourse and Actionable Explanations in Black-Box Decision Making Systems

TL;DR

This work tackles the problem of providing actionable recourse for individuals facing undesirable outcomes in black-box and causal decision systems. It introduces a manifold-aware framework (REVISE) that uses generative models to approximate the data distribution and enables latent-space optimization to produce minimal, realistic changes that flip outcomes while staying on or near the data manifold. The approach extends beyond linear classifiers to causal models, handles immutable attributes, and offers diagnostic capabilities for data bias by revealing how recourse changes interact with confounding. Empirical results on datasets including UCI default Credit, TWINS, and CelebA demonstrate the practicality of manifold-based recourse and its potential to expose bias in decision-making pipelines. Overall, the paper provides a versatile, principled method for generating counterfactual-like recourse that is rooted in the data distribution and applicable to diverse model families.

Abstract

Machine learning based decision making systems are increasingly affecting humans. An individual can suffer an undesirable outcome under such decision making systems (e.g. denied credit) irrespective of whether the decision is fair or accurate. Individual recourse pertains to the problem of providing an actionable set of changes a person can undertake in order to improve their outcome. We propose a recourse algorithm that models the underlying data distribution or manifold. We then provide a mechanism to generate the smallest set of changes that will improve an individual's outcome. This mechanism can be easily used to provide recourse for any differentiable machine learning based decision making system. Further, the resulting algorithm is shown to be applicable to both supervised classification and causal decision making systems. Our work attempts to fill gaps in existing fairness literature that have primarily focused on discovering and/or algorithmically enforcing fairness constraints on decision making systems. This work also provides an alternative approach to generating counterfactual explanations.

Figures (3)

• Figure 1: REVISE Input: ${\mathbf{x}}^* \text{s.t.} \, f({\mathbf{x}}^*) = -1 \\ \mathcal{G}_{\theta}, \mathcal{F}_{\psi}, f, \lambda>0, \eta, \tau_{max}> 0, tt=0$
• Figure 3: (a) Graphical model for a decision making system when all confounders are observed. ${\mathbf{x}}$ is the set of observed attributes (including confounding variables) that affect the binary "treatment" $t$ and the outcome of interest $y$ (b) Graphical model when all confounders are unobserved. (c) Decision making system for which recourse is learned in this work. (d) Recourse under intervention (e, f) Recourse with immutable variables with appropriate intervention.
• Figure 4: Two samples for classifiers $f_2$ (first sub row) and $f_1$ (second sub row) are shown. The leftmost image is the original figure, followed by its reconstruction from the VAE. Intermediate reconstructions are shown as Algorithm \ref{['alg:algo1']} progresses toward crossing the decision boundary (with $\lambda=0.01$). The red bar indicates change in hair color label indicated at the top of each image along with the confidence of prediction. $t$ at the bottom of each image corresponds to the iterations in Algorithm \ref{['alg:algo1']}, shifted by $T$, where $T$ is the iteration where label flips. For both samples, biased classifier $f_2$ shows demonstrable changes in gender specific features ($1^{st}$ and $3^{rd}$ rows) while crossing the decision boundary. At the bottom, we show labels as predicted by a gender classifier for reference.