Recourse Summaries in Automated Decisions

Updated 5 July 2026

Recourse summaries are compact, population-level representations that aggregate individual counterfactual actions into interpretable subgroup prescriptions.
They employ structured forms such as rules and decision trees to enable clear subgroup comparisons and auditing of recourse opportunities.
The approach balances recourse cost, validity, and fairness, ensuring that recommended actions remain feasible, robust, and actionable across demographics.

Searching arXiv for recent and foundational work on recourse summaries. Recourse summaries are compact, population-level representations of actionable recourse in automated decision systems. Whereas individualized recourse specifies a feature change or action set for a single adversely affected instance, recourse summaries aggregate such prescriptions into shared descriptions for subgroups, rules, trees, or other structured objects that support inspection, auditing, and comparison across slices of the population. In the literature, this shift from local to global recourse is motivated by the observation that per-instance counterfactual actions are difficult to aggregate coherently and can obscure disparities in burden, validity, or feasibility across demographic or operational groups (Rawal et al., 2020, Chatzis et al., 8 May 2026). The topic sits at the intersection of algorithmic recourse, counterfactual explanation, fairness auditing, causal inference, and interpretable machine learning, with different formulations emphasizing interpretability, global coverage, cost-effectiveness trade-offs, actionability, and robustness to strategic or causal complications (Karimi et al., 2020, Joshi et al., 2019, Plecko et al., 12 May 2026).

1. Concept and scope

The core distinction in this area is between individualized recourse and global recourse summaries. Individualized recourse asks, for a specific negatively classified point, what minimal feasible action would flip the model’s output. Foundational formulations cast this as a constrained optimization over feasible actions or counterfactual inputs, often under explicit actionability constraints and costs (Ustun et al., 2018, Joshi et al., 2019). In contrast, recourse summaries seek a higher-level description of recourse behavior over a population or subgroup structure, so that one can inspect how recourse is distributed rather than solving a separate optimization problem for each individual (Rawal et al., 2020).

A widely used motivation is auditing. If the objective is to determine whether different groups systematically receive different recourse opportunities or burdens, it is insufficient to inspect isolated counterfactuals. A summary representation makes subgroup-level comparison tractable: it exposes who is affected, what common action is recommended, how costly that action is, and how often it succeeds (Chatzis et al., 8 May 2026, Rawal et al., 2020). This suggests that recourse summaries are not merely compressed explanations; they are instruments for model governance.

The survey literature places this development within a broader conceptual divide between contrastive explanations and consequential recommendations. A contrastive explanation identifies an alternative feature profile that would change the prediction, while a consequential recommendation specifies feasible actions that could lead to such a profile in the real world (Karimi et al., 2020). Recourse summaries may summarize either level, but the stronger forms aim to summarize actionable remedies rather than abstract counterfactuals. This is particularly important because explanation alone does not guarantee real-world feasibility (Karimi et al., 2020).

2. Formal formulations of summary recourse

Several formalizations of recourse summaries appear in the literature, differing mainly in the structure used to partition the affected population and assign actions.

In AReS, or Actionable Recourse Summaries, the explanation is a two-level recourse set $R$ made of triples $(q_i,c_i,c'_i)$ , where $q_i$ identifies a subgroup, $c_i$ is an antecedent, and $c'_i$ is the associated recourse action (Rawal et al., 2020). The affected population is

$\mathcal{X}_{\text{aff}} = \{x \in \mathcal{X} : B(x)=0\}.$

The framework optimizes correctness, coverage, and cost-like objectives under explicit interpretability constraints: $\argmax_{\mathcal{R} \subseteq \mathcal{SD}\times \mathcal{RL}} \sum_{i=1}^{4}\lambda_i f_i(R)$ subject to size, width, and number-of-rule-set limits (Rawal et al., 2020). In this formulation, recourse summaries are explicitly global, interpretable, and model-agnostic.

In SOGAR, or Summaries of Optimal and Global Actionable Recourse, a summary is a decision tree $\tau : \mathcal{X} \to \mathcal{A}$ that maps each adversely classified instance to an action (Chatzis et al., 8 May 2026). The affected population is defined as

$D_0 = \{ x^{(i)} : \langle x^{(i)}, f(x^{(i)}) \rangle \in D, ~f(x^{(i)}) = 0 \},$

and the summary-level objectives are

$C(h) = \sum_{x \in D_0} c(h(x), x), \qquad L(h) = \sum_{x \in D_0} l(h(x), x),$

with

$(q_i,c_i,c'_i)$ 0

The design problem is bi-objective: lower cost and lower loss generally conflict, so SOGAR searches for the Pareto front of non-dominated summaries rather than fixing a scalarization in advance (Chatzis et al., 8 May 2026).

A third line, exemplified by T-CREx, represents recourse summaries as rules and metarules. Counterfactual rules denote regions of the input space associated with favorable outputs, while metarules denote regions where a given rule is optimal under a specified cost (Bewley et al., 2024). This provides both individual lookup and global structure. A counterfactual rule $(q_i,c_i,c'_i)$ 1 is filtered through feasibility and accuracy constraints, and the chosen rule minimizes

$(q_i,c_i,c'_i)$ 2

where sparsity dominates feasibility lexicographically in practice (Bewley et al., 2024).

These formulations all preserve a common pattern: define the negatively affected population, partition it into interpretable regions, assign a shared action or rule per region, and optimize under some combination of validity, cost, plausibility, and interpretability (Rawal et al., 2020, Chatzis et al., 8 May 2026, Bewley et al., 2024).

3. Representation families

Recourse summaries have been instantiated through several representational families, each with distinct strengths.

Rule-based summaries are central in AReS and T-CREx. AReS uses compact nested rule sets, with outer rules describing subpopulations and inner rules specifying recourse (Rawal et al., 2020). T-CREx uses tree-derived hyperrectangular rules and metarules, learned through surrogate models, to summarize both local and global counterfactual structure (Bewley et al., 2024). Rule-based summaries are especially useful when the objective is human-readable auditing or subgroup comparison.

Tree-based summaries are emphasized in SOGAR. Internal nodes use shallow axis-parallel splits of the form

$(q_i,c_i,c'_i)$ 3

and leaves are assigned sparse actions $(q_i,c_i,c'_i)$ 4 (Chatzis et al., 8 May 2026). The shallow-tree constraint limits subgroup complexity, while sparse leaf actions bound the number of feature edits. This combination is explicitly motivated by stability and interpretability (Chatzis et al., 8 May 2026).

Set-of-changes summaries appear in methods that remain individual-first but return concise attribute modifications. REVISE, for example, returns the coordinate-wise difference between the original point and a generated manifold-respecting counterfactual: $(q_i,c_i,c'_i)$ 5 producing a compact actionable explanation for a single individual (Joshi et al., 2019). While not a global summary in itself, this format is one precursor to later summary mechanisms.

Path-based plans appear in diverse-recourse work such as the interpolation-based approach that first selects favorable prototypes and then connects them by feasible paths in an actionability graph (Nguyen et al., 2023). The output is a set of actionable plans rather than a single counterfactual. This is not a global subgroup summary in the AReS or SOGAR sense, but it broadens the notion of a recourse summary to include structured multi-step alternatives.

A plausible implication is that the choice of representation reflects the intended use: rule and tree summaries favor auditing and subgroup diagnosis; path summaries favor end-user choice among feasible plans; manifold-based change sets favor concise individual guidance (Rawal et al., 2020, Chatzis et al., 8 May 2026, Nguyen et al., 2023, Joshi et al., 2019).

4. Objectives: correctness, coverage, cost, and trade-offs

The defining technical problem in recourse summaries is not only finding a valid action, but balancing multiple competing desiderata.

In AReS, the central objectives are recourse correctness, coverage, feature cost, and feature change magnitude (Rawal et al., 2020). The incorrect-recourse term counts affected individuals for whom the prescribed subgroup action does not flip the black-box prediction; coverage counts how many affected individuals are described by some rule; feature cost and change magnitude encode actionability and burden. Interpretability is enforced through constraints on rule-set size, maximum width, and number of distinct subgroup descriptors (Rawal et al., 2020).

In SOGAR, the trade-off is framed more explicitly as recourse cost versus recourse loss (Chatzis et al., 8 May 2026). Cost is measured in the main experiments by Maximum Percentile Shift,

$(q_i,c_i,c'_i)$ 6

and loss by failure-to-flip,

$(q_i,c_i,c'_i)$ 7

Rather than collapsing these objectives into a weighted sum, SOGAR computes the Pareto front

$(q_i,c_i,c'_i)$ 8

so that improving one objective necessarily worsens the other (Chatzis et al., 8 May 2026). This is especially relevant for auditing, because disparities may vary across the trade-off surface rather than at a single operating point.

Rule-based approaches such as T-CREx emphasize accuracy, feasibility, sparsity, complexity, consistency, and runtime (Bewley et al., 2024). There, the method does not directly optimize pointwise recourse distance, but the rule structure indirectly affects distance through feasible-region geometry. This differs from AReS and SOGAR, where aggregate population cost enters the core objective.

More generally, the survey literature distinguishes distance from cost. Distance compares feature profiles, whereas cost measures real-world effort, and the two coincide only under restrictive assumptions (Karimi et al., 2020). This distinction matters acutely for summaries: a concise global rule may describe a small feature displacement while still masking large implementation effort, asymmetric effort across directions, or latent causal dependencies (Karimi et al., 2020).

5. Interpretability and auditing

A principal reason to build recourse summaries is to make recourse behavior inspectable before deployment. This emphasis is explicit in AReS, which argues that it is not enough to generate individual recourses; one must analyze and interpret a predictive model and vet its recourses to ensure that they are meaningful and non-discriminatory (Rawal et al., 2020).

AReS supports both automatically discovered and user-specified subgroup descriptors, enabling targeted questions such as whether recourses differ across protected or operationally relevant groups (Rawal et al., 2020). T-CREx similarly supports a global analysis of model behavior by cataloging which recourse rules exist and where they are optimal (Bewley et al., 2024). SOGAR operationalizes auditing through shared subgroup actions defined by a shallow tree, which makes it straightforward to compare subgroup-average cost and loss across the entire Pareto front (Chatzis et al., 8 May 2026).

The auditing use case is tightly connected to fairness-of-recourse work. “Equalizing Recourse across Groups” defines recourse as the minimum distance to a decision boundary and measures disparity through the gap in average recourse over negatively classified members of different groups: $(q_i,c_i,c'_i)$ 9 with a training objective that constrains this gap (Gupta et al., 2019). Although that work is not itself about global summary construction, it supplies one of the main normative motivations: unequal post-decision improvement opportunities constitute a practical form of unfairness (Gupta et al., 2019).

Ranked settings extend the same logic. RAGUEL introduces ranked group-level recourse fairness, requiring not only fair representation across ranking prefixes but also balance in mean recourse cost across groups (Haldar et al., 2022). This widens the summary idea beyond binary classification to ranked outputs, where subgroup-level recourse burden can be summarized at every prefix or block of the ranking (Haldar et al., 2022).

A common misconception is that standard fairness constraints on outcomes automatically imply fair recourse. The literature repeatedly rejects this. Models can satisfy output-based fairness notions while imposing very different recourse costs across groups, and post-processing methods aimed at demographic parity or equalized odds may even exacerbate recourse unfairness (Gupta et al., 2019, Haldar et al., 2022).

6. Causality, realism, and robustness

A persistent challenge for recourse summaries is that global compactness can conceal whether the recommended actions are realistic, causally meaningful, or robust.

Realism is central in REVISE, which seeks the smallest plausible set of feature changes by moving in the latent space of a generative model rather than performing arbitrary off-manifold perturbations (Joshi et al., 2019). The underlying constrained problem is

$q_i$ 0

and in practice the search is performed in latent space with a generator $q_i$ 1 (Joshi et al., 2019). This directly addresses a limitation of purely geometric summaries: compact prescriptions may be invalid as real states of the world.

Causality sharpens the problem. Improvement-Focused Causal Recourse distinguishes acceptance by the predictor from improvement in the underlying target and argues that recourse aimed solely at acceptance may amount to gaming the predictor (König et al., 2022). Its objective is to find low-cost interventions satisfying an improvement-confidence threshold,

$q_i$ 2

with $q_i$ 3 defined in individualized or subpopulation-based terms (König et al., 2022). This suggests that a subgroup-level recourse summary can be misleading if it summarizes actions that flip predictions without improving the real-world outcome.

The 2026 causal foundations paper generalizes this by arguing that recourse should be modeled as a repeated-decision process over pre- and post-intervention outcomes, not as a frozen counterfactual of a single unit (Plecko et al., 12 May 2026). Under post-recourse stability assumptions, observational data may identify recourse effects; when paired recourse data exist, copula parameters can be estimated and tested (Plecko et al., 12 May 2026). A plausible implication is that future summary methods may need to summarize not only recommended actions, but also the uncertainty structure connecting pre- and post-recourse states.

Robustness to model updates is another concern. ROAR shows that recourses optimized against a fixed model can be invalidated after retraining and proposes a minimax objective over plausible model shifts: $q_i$ 4 The result is a shift-aware form of recourse that remains valid under a specified perturbation family (Upadhyay et al., 2021). Summary methods rarely encode such robustness directly, so a global summary may become stale if the deployed model changes.

Risk-aware sequential recourse adds a related dimension. SafeAR treats recourse as a finite-horizon MDP and summarizes the distribution of recourse costs through success rate, mean/variance, Value at Risk, and Conditional Value at Risk (Wu et al., 2023). This broadens what “summary” can mean: not only what action is recommended, but what downside-tail burden it induces under uncertainty.

Recourse summaries also expose tensions that are not purely algorithmic. Several papers emphasize that recourse is embedded in strategic interaction, institutional incentives, and benchmarking infrastructure.

“Incentives to Offer Algorithmic Recourse” shows that a decision-maker offers recourse only for an intermediate range of applicant probabilities,

$q_i$ 5

where $q_i$ 6 is continuation probability for profitable applicants and $q_i$ 7 is manipulability by unprofitable ones (Olckers et al., 2023). “To Give or Not to Give? The Impacts of Strategically Withheld Recourse” extends this logic by showing that a utility-maximizing system may strategically reveal only a subset of recourse actions because disclosure leaks information that enables imitation and false positives (Chen et al., 8 Apr 2025). In that setting, the system solves

$q_i$ 8

anticipating strategic response (Chen et al., 8 Apr 2025). These results imply that any global summary of available recourse may itself be shaped by institutional incentives rather than by neutral explanation design.

Population-level harm is also now treated in broader modalities. Collective recourse for generative urban visualization systems reframes recourse as a community-level reporting and mitigation process rather than as individual feature change (Mushkani, 15 Sep 2025). Although this lies outside binary classification, it preserves the summary logic: harms are aggregated, triaged, and addressed through structured collective reports rather than one-off personalized fixes.

Finally, reproducibility has become an explicit part of the field’s infrastructure. RecourseBench decomposes evaluation into Data, Preprocessing, Model, Recourse Method, and Evaluation layers and validates integrated methods against originally reported results through an automated test suite (Khotanlou et al., 15 Jun 2026). The benchmark includes 28 recourse methods and defines standardized metrics such as validity,

$q_i$ 9

proximity, sparsity, plausibility, and runtime (Khotanlou et al., 15 Jun 2026). This matters for recourse summaries because global methods are especially sensitive to implementation details, subgroup definitions, and metric choices; without reproducible evaluation, cross-method comparisons can be misleading.

Taken together, these developments show that recourse summaries have evolved from compact descriptive devices into a broader framework for auditing, fairness analysis, causal reasoning, and governance. The field has moved from summarizing minimal feature edits for individuals (Joshi et al., 2019, Ustun et al., 2018) toward learning structured population-level objects that expose recourse trade-offs and subgroup disparities (Rawal et al., 2020, Chatzis et al., 8 May 2026, Bewley et al., 2024). At the same time, causal, robust, and strategic perspectives have made clear that a summary is only as meaningful as the assumptions under which the underlying recourse is valid (König et al., 2022, Upadhyay et al., 2021, Olckers et al., 2023, Chen et al., 8 Apr 2025).