Recourse Training Regime

• Recourse training regimes are methods that embed the recourse objective directly into model training to guarantee feasible, low-cost changes for altering model outcomes.
• They leverage modified optimization objectives—such as inverted adversarial and risk-aware objectives—to simultaneously address feasibility, plausibility, and robustness under model or data shift.
• Advanced approaches integrate sequential decision policies and personalization to generate recourses that remain valid and safe in dynamic, high-stakes environments.

Searching arXiv for recent recourse-training papers and related formulations to ground the article. arxiv_search({"query":"all:(algorithmic recourse training regime OR recourse-aware training OR robust recourse)","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}) Recourse, in the foundational formulation of linear classification, is the ability of a person to change the decision of a model by altering actionable input variables, such as income rather than immutable features like age or marital status (Ustun et al., 2018). The core optimization problem asks for a feasible action $a$ of minimum cost such that the fixed predictor changes from an unfavorable to a favorable outcome: $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$ A plausible unifying interpretation of a “recourse training regime” is therefore an Editor’s term for any learning or optimization procedure that internalizes this recourse objective during model fitting, recourse-generator training, or policy synthesis, rather than treating recourse as a purely post-hoc search problem. Across the literature, such regimes differ in whether they optimize for existence of recourse, low-cost recourse, plausible recourse, robust recourse under model shift, personalized recourse, or risk-aware sequential recourse (Ross et al., 2020, Guo et al., 2022, Garg et al., 12 May 2025).

1. Foundational formulation and the move beyond post-hoc search

The early linear-classification formulation established two elements that remain standard in later work: a distinction between actionable and immutable features, and a feasibility notion in which recourse is absent when the constrained optimization problem has no solution (Ustun et al., 2018). For linear classifiers $f(x)=\operatorname{sign}(w^\top x)$, the same work introduced integer programming tools to check recourse feasibility and cost, and to enumerate minimal actionable change-sets, or “flipsets,” without interfering in model development (Ustun et al., 2018). It also showed experimentally that feature selection, regularization, and deployment on unrepresentative populations can substantially alter whether recourse exists and how costly it is (Ustun et al., 2018).

This post-hoc paradigm motivated a later shift toward training-time intervention. “Learning Models for Actionable Recourse” framed the central deficiency of post-hoc methods as the absence of guarantees that actionable recourses exist for most affected individuals, and proposed to train the predictive model itself so that recourse exists with high probability (Ross et al., 2020). The key modification is an “inverted” adversarial objective: $$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$ which differs from conventional adversarial training by using $\min_{\delta\in\Delta}$ and by forcing the target label to $1$ (Ross et al., 2020). This is coupled with PAC confidence sets to produce the PARE guarantee: with probability at least $1-\alpha$ over the calibration sample, at least $1-\epsilon$ of future negative individuals will be provided with valid recourse (Ross et al., 2020).

A related but distinct development appeared for tree-based models. “Learning Decision Trees and Forests with Algorithmic Recourse” formulated learning under an explicit recourse-risk constraint, where the empirical objective minimizes prediction risk subject to a bound on the fraction of points without valid recourse under a cost budget $\varepsilon$ (Kanamori et al., 2024). The proposed Recourse-Aware Classification Tree uses a penalized split criterion,

$$\Phi_\lambda(d,b,y_L,y_R)=\hat{R}(h' \mid S)+\lambda\cdot \hat{\Omega}_\varepsilon(h' \mid S),$$

and a post-hoc weighted partial set-cover relabeling step to ensure recourse for as many instances as possible while keeping accuracy degradation minimal (Kanamori et al., 2024). This suggests that recourse training regimes are not tied to differentiable models; they can also be built into greedy, combinatorial learners.

2. Joint objectives: proximity, plausibility, validity, and guaranteed feasibility

A recurring theme is that effective recourse must satisfy multiple objectives simultaneously. “From Search To Sampling: Generative Models For Robust Algorithmic Recourse” makes this explicit by identifying three conflicting goals: proximity to the original profile, plausibility for realistic recourse, and validity to ensure the desired outcome (Garg et al., 12 May 2025). The paper argues that existing methods often train for these objectives separately and only combine them at inference time through joint optimization, frequently via non-robust gradient descent search (Garg et al., 12 May 2025). GenRe instead defines a conditional generative model $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$0 and trains it end-to-end against an ideal recourse distribution

$$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$1

where $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$2 models proximity and immutability constraints, $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$3 models plausibility, and $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$4 encodes validity (Garg et al., 12 May 2025). Because direct recourse supervision is unavailable, GenRe synthesizes supervision through a proximity-weighted pairing distribution over highly confident positive examples and proves that this target is a consistent estimator of the ideal recourse distribution (Garg et al., 12 May 2025).

The contrast with the linear and integer-programming tradition is instructive. In the linear setting, feasibility and minimality are certified directly by optimization over $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$5, including custom non-linear or non-convex cost functions such as maximum percentile shift

$$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$6

and log-ratio costs for flipset enumeration (Ustun et al., 2018). In the generative setting, the same desiderata are encoded statistically in the conditional model and recovered by forward sampling rather than per-instance optimization (Garg et al., 12 May 2025).

The literature therefore supports two training-level strategies. One strategy trains the predictive model so that recourse is likely or guaranteed to exist, as in PARE and recourse-aware tree induction (Ross et al., 2020, Kanamori et al., 2024). The other trains a dedicated recourse model so that returned counterfactuals already internalize multiple recourse objectives, as in GenRe (Garg et al., 12 May 2025). A plausible implication is that “recourse training regime” refers not to a single architecture class but to the placement of recourse desiderata inside the learning objective.

3. Robustness-oriented regimes under model and data shift

A major line of work treats recourse as unreliable unless it remains valid after the predictive model changes. “Towards Robust and Reliable Algorithmic Recourse” observed that deployed models are regularly updated and proposed RObust Algorithmic Recourse (ROAR), which replaces the standard recourse objective with a minimax objective over plausible model shifts: $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$7 For linear and locally linearized models, ROAR uses an adversarial training-inspired inner maximization over model perturbations to find recourses that remain valid under shift, and provides both a lower bound on invalidation probability for non-robust recourses and an upper bound on the additional cost incurred by robust recourse (Upadhyay et al., 2021).

RoCourseNet extends this robustness perspective from recourse optimization to joint training of prediction and recourse generation (Guo et al., 2022). It formulates robust recourse generation as a tri-level problem: an outer minimization over predictor and recourse-generator parameters, a middle maximization over worst-case data shifts, and an inner minimization that retrains the predictor on shifted data. The Virtual Data Shift algorithm approximates the worst-case shift by unrolling gradient descent steps on the shifted model and then taking projected gradient ascent steps in data space, while block-wise coordinate descent alternates predictor updates with recourse-generator updates (Guo et al., 2022). The reported result is more than 96% robust validity on three real-world datasets and at least a 10% improvement over state-of-the-art baselines in robust counterfactual generation (Guo et al., 2022).

A different robustness question concerns whether adversarially robust classifiers necessarily make recourse more expensive. “Adaptive Adversarial Training Does Not Increase Recourse Costs” studies instance-wise adaptive adversarial radii rather than fixed-radius adversarial training (Hardy et al., 2023). Traditional adversarial training uses

$$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$8

whereas adaptive adversarial training uses instance-specific radii $$\min \operatorname{cost}(a,x)\quad \text{s.t.}\quad f(x+a)=+1,\; a\in A(x).$$9 (Hardy et al., 2023). The paper’s finding is that AAT and MMA achieve notable robustness improvements with little effect on recourse costs, preserving a much larger proportion of low-cost recourse than fixed-radius training and often keeping average recourse costs close to those of naturally trained models (Hardy et al., 2023). This directly rebuts the common assumption that robustness and affordable recourse must always be in tension.

4. Sequential, personalized, and risk-aware policy learning

Several works generalize recourse from single-point counterfactuals to sequential decision problems. “Personalized Algorithmic Recourse with Preference Elicitation” treats algorithmic recourse as a sequence of actions whose cost depends on a user-specific latent preference vector $f(x)=\operatorname{sign}(w^\top x)$0 (Toni et al., 2022). PEAR maintains a posterior $f(x)=\operatorname{sign}(w^\top x)$1, asks choice-set queries chosen by maximizing Expected Utility of Selection,

$f(x)=\operatorname{sign}(w^\top x)$2

and integrates the elicitation loop into a reinforcement-learning agent coupled with Monte Carlo Tree Search (Toni et al., 2022). The interaction rounds refine the posterior over user costs, and empirical results show high-quality personalized recourse in only a handful of iterations (Toni et al., 2022).

“Personalized Path Recourse for Reinforcement Learning Agents” moves fully into trajectory space (Hong et al., 2023). Given an original path $f(x)=\operatorname{sign}(w^\top x)$3, a desired goal $f(x)=\operatorname{sign}(w^\top x)$4, and an agent policy $f(x)=\operatorname{sign}(w^\top x)$5, it trains a recourse agent $f(x)=\operatorname{sign}(w^\top x)$6 with a composite reward

$f(x)=\operatorname{sign}(w^\top x)$7

where similarity is based on Levenshtein distance and personalization is a trajectory-level sum of a custom link function applied to policy likelihoods (Hong et al., 2023). The recourse agent is trained with DQN, experience replay, target networks, and a UCB-inspired exploration function, and experiments in Gridworld, Super Mario Bros, text recourse, and temperature sequences show that the method adapts recourse to different agents’ behavior while trading off goal satisfaction, similarity, and personalization (Hong et al., 2023).

Risk-aware sequential recourse appears in SafeAR. This framework models recourse as a finite-horizon MDP and optimizes not only expected total cost but also its variability, with value function

$f(x)=\operatorname{sign}(w^\top x)$8

and action-value recursion

$f(x)=\operatorname{sign}(w^\top x)$9

SafeAR evaluates policies using Value at Risk and Conditional Value at Risk,

$$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$0

and shows that risk-averse policies reduce variance, VaR, and CVaR at the cost of slightly higher mean cost (Wu et al., 2023). This broadens the meaning of recourse training regime from “make recourse exist” to “make recourse safe under uncertainty.”

5. Generative, black-box, causal, and domain-specific extensions

In domain-specific settings, recourse training regimes have been implemented through specialized generative or black-box procedures. “Actionable Recourse via GANs for Mobile Health” uses CounteRGAN, with a generator $$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$1, discriminator $$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$2, and fixed classifier $$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$3, to learn residual interventions $$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$4 such that

$$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$5

The generator and discriminator are trained adversarially in alternation, training uses only users expected to churn in fewer than 90 days, and checkpointing is performed when discriminator accuracy is near random, indicating realism of generated counterfactuals (Chien et al., 2022). The paper reports that GAN-generated recourse can outperform gradient-descent-based alternatives in percent successful recourse while being much faster at inference (Chien et al., 2022).

For large black-box tabular predictors, “Algorithmic Recourse of In-Context Learning for Tabular Data” studies in-context learning rather than conventional training (Dong et al., 29 May 2026). The paper proves that recourse remains well-defined and bounded under ICL, gives a high-probability upper bound on optimal recourse cost, and shows that as context size grows, recourse converges toward classical fixed-model solutions (Dong et al., 29 May 2026). Its practical algorithm, Adaptive Subspace Recourse for In-Context Learning, solves

$$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$6

through zeroth-order optimization over adaptively selected feature subspaces, enforcing feasibility constraints by projection and supporting multi-class tabular tasks (Dong et al., 29 May 2026). Although this is not a training regime for the predictor, it is a recourse regime for black-box inference-time models.

The causal perspective further shifts attention from static counterfactuals to observed post-intervention outcomes. “Causal Algorithmic Recourse: Foundations and Methods” models recourse as a joint process over pre- and post-recourse outcomes in a structural causal model with partially resampled latent variables (Plecko et al., 12 May 2026). Under Post-Recourse Stability, some recourse effects can be inferred from observational data alone; with paired recourse data, the paper estimates copula parameters by conditional maximum likelihood,

$$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$7

tests goodness-of-fit, and switches to a distribution-free regression method when the copula model is rejected (Plecko et al., 12 May 2026). This yields a staged regime in which recourse recommendations are updated as recourse outcome data accumulate.

6. Evaluation criteria, trade-offs, and criticisms

The literature evaluates recourse regimes along several axes: feasibility or validity, cost or proximity, plausibility or manifold adherence, sparsity, personalization, robustness under model shift, query efficiency, and, increasingly, privacy and population-level risk (Ustun et al., 2018, Garg et al., 12 May 2025, Hardy et al., 2023, Huang et al., 2023). The following summary captures recurring emphases.

Regime family	Primary training or optimization signal	Representative papers
Predictive-model training	Existence of feasible recourse, empirical recourse risk, PAC guarantees	(Ross et al., 2020, Kanamori et al., 2024)
Robust recourse training	Worst-case model or data shift, robust validity	(Upadhyay et al., 2021, Guo et al., 2022)
Sequential/personalized policy learning	Goal achievement, similarity, personalization, inferred user cost, risk sensitivity	(Toni et al., 2022, Hong et al., 2023, Wu et al., 2023)
Generative recourse models	Joint proximity–plausibility–validity likelihood	(Garg et al., 12 May 2025, Chien et al., 2022)

One controversy is whether recourse is beneficial at all. “The Risks of Recourse in Binary Classification” introduces a learning-theoretic framework comparing expected loss with and without recourse,

$$\ell_R(\theta)=\mathbb{E}_{p(x,y)}\left[\ell(g_\theta(x),y)+\lambda\cdot \min_{\delta\in\Delta}\ell(g_\theta(x+\delta),1)\right],$$8

and shows that for Bayes-optimal or near-Bayes-optimal classifiers, recourse can increase population-level risk by moving users toward regions of higher class uncertainty (Fokkema et al., 2023). The paper also shows that deployers may have incentives to strategize in anticipation of having to provide recourse, sometimes to the detriment of users (Fokkema et al., 2023). This is a fundamental critique of recourse regimes that optimize only label flipping.

A second controversy concerns privacy. “Accurate, Explainable, and Private Models: Providing Recourse While Minimizing Training Data Leakage” shows that recourse can leak membership information through counterfactual distance and proposes two differentially private mitigations: a differentially private model and Laplace Recourse (Huang et al., 2023). In the latter, noise is added directly to the predicted probability before converting to a noisy logit and computing the counterfactual distance. The paper reports that both methods reduce what an adversary can infer, especially at low false-positive rate, with Laplace Recourse often preserving model accuracy because it does not alter training (Huang et al., 2023).

A third persistent issue is distributional and subgroup disparity. The foundational linear work showed that recourse costs can differ between protected groups even when the protected attribute is not used by the classifier, because correlated features affect the recourse solution (Ustun et al., 2018). It also showed that out-of-sample deployment and biased training samples can sharply increase cost or completely remove recourse for certain subpopulations (Ustun et al., 2018). This suggests that any recourse training regime must be evaluated not only for average validity or cost but also for whose recourse remains feasible after deployment.

Taken together, the literature supports a broad but technically coherent view of recourse training regimes. They are procedures that move recourse from an after-the-fact explanation problem into the training or optimization objective itself: by guaranteeing recourse existence, synthesizing robust or personalized recourse policies, learning conditional recourse distributions, or grounding recourse in observed post-intervention data (Ross et al., 2020, Guo et al., 2022, Garg et al., 12 May 2025, Plecko et al., 12 May 2026). At the same time, the literature also shows that better recourse training is not equivalent to beneficial recourse: robustness, privacy, fairness, and population-level risk remain independent constraints on whether a recourse regime is trustworthy in high-stakes deployment (Hardy et al., 2023, Huang et al., 2023, Fokkema et al., 2023).