Improvement-Focused Causal Recourse (ICR)

TL;DR

This work introduces Improvement-Focused Causal Recourse (ICR), a paradigm that shifts recourse from merely achieving acceptance of the predictor to actively improving the underlying target $Y$ by leveraging causal knowledge. It defines two improvement confidences, $\\gamma^{ind}$ and $\\gamma^{sub}$, and develops an optimization framework to discover cost-efficient actions that reliably improve $Y$ while ensuring post-recourse decisions translate into acceptance via carefully designed post-recourse predictors. The approach yields both improvement and acceptance guarantees under correct causal modeling, and empirical results on synthetic and semi-synthetic data show ICR outperforms traditional CE/CR in producing meaningful real-world improvements and robust acceptance, even under model refits. The paper also discusses limitations related to causal knowledge requirements, identifiability, and contestability, highlighting practical implications for deploying causally informed recourse in high-stakes settings.

Abstract

Algorithmic recourse recommendations, such as Karimi et al.'s (2021) causal recourse (CR), inform stakeholders of how to act to revert unfavourable decisions. However, some actions lead to acceptance (i.e., revert the model's decision) but do not lead to improvement (i.e., may not revert the underlying real-world state). To recommend such actions is to recommend fooling the predictor. We introduce a novel method, Improvement-Focused Causal Recourse (ICR), which involves a conceptual shift: Firstly, we require ICR recommendations to guide towards improvement. Secondly, we do not tailor the recommendations to be accepted by a specific predictor. Instead, we leverage causal knowledge to design decision systems that predict accurately pre- and post-recourse. As a result, improvement guarantees translate into acceptance guarantees. We demonstrate that given correct causal knowledge, ICR, in contrast to existing approaches, guides towards both acceptance and improvement.

Figures (9)

• Figure 1: Directed Acyclic Graph (DAG) illustrating the perspective on model and data taken by counterfactual explanations (CE, left) and causal recourse (CR, center) in contrast to improvement-focused recourse (ICR, right). Blue edges represent the causal links induced by the prediction model, green edges the real-world causal links, gray nodes the covariates, and the red (yellow) node the primary (secondary) recourse target. CR respects the causal relationships but only between input features. ICR is the only approach that takes the target $Y$ into account. While CE and CR aim to revert the prediction $\hat{Y}$, ICR aims to revert the target $Y$.
• Figure 2: Experimental results for CE, CR and ICR on four datasets over $10$ runs on $200$ individuals each. For the probabilistic methods the confidences $0.75, 0.85, 0.9, 0.95$ were targeted (for CR: $\overline{\eta}$, for ICR: $\overline{\gamma}$). For CE no slack is allowed, such that the results correspond to a confidence level of $1.0$. Values are reported on a quadratic scale.
• Figure 3: Causal graph $\mathcal{G}_{\overline{I_a}}$ visualizing the subpopulation-based post-recourse setting, including the prediction target $Y$ (light blue), intervened-upon variables $I_a$ (red), the subgroup characteristics $G_a$ (cyan) and the descendants $\Gamma$ that shall be resampled (dark blue). $\overline{I_a}$ indicates that incoming edges to $I_a$ were removed. Right: Causal graph $\mathcal{G}_{\overline{I_a}\underline{G_a}}$ where incoming edges to $I_a$ and outgoing edges from $G_a$ were removed. We observe that in this manipulated graph $G_a$ is $d$-separated from $\Gamma$. Thus, according to the second rule of $do$-calculus, for $G_a$ intervention and conditioning coincide.
• Figure 4: SCM for 3var-causal. The cost function is given as $cost(a) = \delta_1 + \delta_2 + \delta_3$, where $\delta$ is the vector of absolute changes to the intervened upon variables. E.g., for $do(a) = do(X_1=x_1')$, $\delta_1 = |x_1' - x_1|$ and $\delta_2 = \delta_3 = 0$
• Figure 5: SCM for 3var-noncausal with $cost(a) = \delta_1 + \delta_2 + \delta_3$.

Theorems & Definitions (11)

• Definition 1: Individualized improvement confidence
• Definition 2: Subpopulation-based improvement confidence
• Definition 3: Individualized post-recourse predictor
• Proposition 1
• Proposition 2
• Proposition 3
• Proposition 4
• Proposition 5
• Proposition 6
• Example 1