From Universal to Individualized Actionability: Revisiting Personalization in Algorithmic Recourse
Authors: Lena Marie Budde, Ayan Majumdar, Richard Uth, Markus Langer, Isabel Valera
Organizations: Saarland Informatics Campus, Saarland University, Saarbrücken, Germany · Max Planck Institute for Software Systems, Saarbrücken, Germany · Institute for Psychology, University of Freiburg, Freiburg, Germany
Algorithmic recourse aims to provide actionable recommendations that enable individuals to change unfavorable model outcomes, and prior work has extensively studied properties such as efficiency, robustness, and fairness. However, the role of personalization in recourse remains largely implicit and underexplored. While existing approaches incorporate elements of personalization through user interactions, they typically lack an explicit definition of personalization and do not systematically analyze its downstream effects on other recourse desiderata. In this paper, we formalize personalization as individual actionability, characterized along two dimensions: hard constraints that specify which features are individually actionable, and soft, individualized constraints that capture preferences over action values and costs. We operationalize these dimensions within the causal algorithmic recourse framework, adopting a pre-hoc user-prompting approach in which individuals express preferences via rankings or scores prior to the generation of any recourse recommendation. Through extensive empirical evaluation, we investigate how personalization interacts with key recourse desiderata, including validity, cost, and plausibility. Our results highlight important trade-offs: individual actionability constraints, particularly hard ones, can substantially degrade the plausibility and validity of recourse recommendations across amortized and non-amortized approaches. Notably, we also find that incorporating individual actionability can reveal disparities in the cost and plausibility of recourse actions across socio-demographic groups. These findings underscore the need for principled definitions, careful operationalization, and rigorous evaluation of personalization in algorithmic recourse.
Figures & tables
Figure 1: iCARMA architecture (adapted from [ 34 ] ); red parts indicate modifications to enable individualized recommendations.
Table 1: Impact of hard actionability constraints on recourse metrics for amortized iCARMA and non-amortized Oracle.
Figure 3: Distribution of the log densities for the factual and counterfactual (cf) under different numbers of actionable features (indicated in brackets).
dataset+constraints
individual cost profile
method
validity
plausibility
cost
hap
Loan, soft
constant
oracle
0.959
1.0
0.057±0.064
0.614
iCARMA
0.999
0.886
0.069±0.077
0.852
concave
oracle
0.959
1.0
0.074±0.088
0.492
iCARMA
0.998
0.631
0.079±0.088
0.044
linear
oracle
0.959
1.0
0.072±0.085
0.443
iCARMA
0.996
0.532
0.077±0.086
0.029
Table 2: Impact of preference consideration as rankings on recourse metrics across four cost profiles: concave, linear, convex, and constant (no cost individualization) for amortized iCARMA and non-amortized causal Oracle.
constraint
individual cost profile
validity
plausibility
cost
hap
soft
constant
0.996
0.841
0.067±0.07
0.829
concave
0.995
0.713
0.073±0.078
0.225
linear
0.991
0.646
0.074±0.086
0.178
convex
0.982
0.5
0.081±0.082
0.194
hard (random) + soft
constant
0.991
0.712
0.073±0.083
0.838
concave
0.99
0.618
0.076±0.087
0.415
Table 3: Impact of preference consideration as scores on recourse metrics across cost profiles for iCARMA on the Loan Dataset.
Figure 4: Cost (left) and distributional log-density (right) results of recourse recommendations across intersectional gender and race groups for four scenarios: (1) non-personalized: all globally actionable features are equally actionable, (2) random-assigned: privileged preference profiles assigned randomly, (3) race-correlated: 90% of white, 5% of non-white individuals assigned the privileged profile, (4) income-dependent: top-earning 25% of individuals assigned the privileged profile.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
prefu,i=1
prefu,i=2
prefu,i=3
prefu,i=4
uniform
co
rco
privileged pref profile
non-privileged pref profile
LA
Dur
Inc
Sav
1/24
1/24
0
0
1/6
LA
Dur
Sav
Inc
1/24
1/24
0
0
1/6
LA
Inc
Dur
Sav
1/24
1/6
0
0
1/24
LA
Sav
Dur
Inc
1/24
1/24
0
0
1/24
LA
Inc
Sav
Dur
1/24
1/6
0
0
1/24
LA
Sav
Inc
Dur
1/24
1/24
0
0
1/24
Appendix
Table 4: Different distributions used for preference sampling on the Loan dataset. Each row reports the probability of sampling a particular assignment of preference scores prefu,i to the globally actionable features LA, Dur, Inc, and Sav under a given configuration. The distribution biased according to causal order is denoted by co, while the distribution biased according to reverse causal order is denoted by rco. The distributions used to sample the privileged and non-privileged preference profiles for the fairness case study are reported in the last two columns.
Figure 5: Causal Graph of Loan dataset
Figure 6: Causal graph for Give Me Some Credit. Solid gray nodes indicate globally immutable, light gray shaded nodes indicate mutable but non-actionable, and white nodes indicate globally actionable nodes.
Figure 7: Feature reconstruction for Give Me Some Credit by the best performing CNF model.
bias
number of indivdually actionable features
method
validity
plausibility
cost
causal order aligned
all (4)
oracle
0.959
1.0
0.057_{\text{\pm 0.064}}
iCARMA
0.998
0.864
0.065_{\text{\pm 0.075}}
3
oracle
0.821
1.0
0.066_{\text{\pm 0.078}}
iCARMA
0.997
0.713
0.062_{\text{\pm 0.073}}
2
oracle
0.597
1.0
0.072_{\text{\pm 0.088}}
iCARMA
0.997
0.564
0.066_{\text{\pm 0.09}}
Appendix
Table 5: Impact on recourse metrics for an increasing number of individually actionable features for the hard-constrained case, assuming non-uniform sampling of individually actionable features : we consider a distribution favouring features high in causal order and a distribution favouring features low in causal order (reverse causal order, rco). For the second case, we additionally compare iCARMA trained with individually actionable features sampled uniformly at random with iCARMA trained with individually actionable features sampled from the distribution biased towards features low in causal order. Oracle knows true SCM and uses grid search with hard constraint for plausibility.
distribution
individually actionable
plausibility constraint
validity
plausibility
cost
uniform random
1
No
0.876
0.25
0.135_{\text{\pm 0.151}}
Yes
0.22
1.0
0.077_{\text{\pm 0.096}}
2
No
0.994
0.25
0.096_{\text{\pm 0.133}}
Yes
0.539
1.0
0.078_{\text{\pm 0.096}}
causal order
1
No
0.923
0.28
0.119_{\text{\pm 0.139}}
Yes
0.259
1.0
0.07_{\text{\pm 0.087}}
Appendix
Table 6: Ablation of the plausibility constraint in causal-oracle recourse optimization. Removing the plausibility constraint increases validity but produces less distributionally plausible counterfactuals. The effects of individual actionability preferences on other recourse criteria persist even without plausibility constraints.
Figure 8: Log-density distribution (iCARMA) for Give Me Some Credit and different personalization setups.
Algorithmic recourse provides individuals who have received undesirable outcomes from machine learning models with suggestions for minimum-cost improvements to achieve the desired outcome. A central assumption when computing recourse is that the decision rule remains fixed throughout the recourse implementation phase. We challenge this assumption in settings where individuals compete for limited resources. In such settings, widespread recourse implementation can change the acceptance threshold even when the scoring model that is used to evaluate individuals remains the same. This change in acceptance threshold can, in turn, invalidate the original recourse recommendations (i.e., following the recourse may not lead to the desired outcome). To address this problem, we introduce a framework called recourse under competition that jointly optimizes for recommendation recipients and the recommended score target they need to satisfy to balance the recourse cost and post-shift validity among initially rejected individuals. We develop an algorithm based on the Implicit Function Theorem and empirically analyze its performance. Experiments on synthetic and real datasets show that personalized score targets can achieve higher validity, albeit at a higher cost. In contrast, common score targets generally offer favorable cost-validity trade-offs for lower to medium validity values.
Algorithmic recourse seeks to help individuals reverse unfavorable automated decisions by recommending actionable changes that achieve a desired outcome. As an individual usually has several distinct routes to a favorable decision, and different people can act on different ones, a recourse system should offer multiple realistic alternatives rather than one. Existing approaches formulate recourse as an optimization problem that constructs one or a small set of counterfactuals rather than modeling the underlying space of feasible solutions, and in practice each sacrifices diversity, plausibility, or feasibility to secure the others. We propose Tractable Recourse Distributions, a probabilistic framework that represents the space of feasible alternatives for a given factual instance as a probability distribution over favorable outcomes. For commonly used cost functions based on proximity and the number of feature changes, we show that this distribution admits an exact representation as a probabilistic circuit, obtained by exponentially tilting the circuit; each individual's distribution is therefore available in closed form, without retraining the model. Sampling from these distributions naturally produces diverse and plausible recourses, while the tilting parameters provide explicit control over their proximity and sparsity. Experiments on standard algorithmic recourse benchmark datasets demonstrate that the proposed framework attains diversity, plausibility, and feasibility simultaneously, while retaining sufficient probability mass over feasible counterfactuals for rejection sampling to be practical. A visual study on MNIST illustrates how the tilt strength trades proximity against validity.
Anagha Sabu, Hrithik Suresh, Narayanan C. Krishnan
Mehta Family School of Data Science and Artificial Intelligence Indian Institute of Technology Palakkad Palakkad, Kerala, India
Algorithmic recourse addresses the challenge of providing tailored recommendations to users affected by unfavorable machine learning decisions, in potentially high-stakes scenarios. Traditional approaches to recourse often rely on the closest counterfactual explanations or assume a priori knowledge of a user's causal structure, resulting in interventions that overlook individual contexts and specific feature interactions. To overcome these limitations, we study a human-in-the-loop framework that iteratively approximates the user's structural causal model through interactive queries via Bayesian inference before producing recourse recommendations. This framework exploits humans' feedback to improve the identification of causal effects, allowing personalized recourse that is plausible, cost-effective, and aligned with the actual causal dependencies of each user. As a proof of concept, we evaluate this framework through simulated human responses. Our simulations across linear and non-linear causal models show promising results, though challenges remain in capturing complex, non-linear structures, emphasizing the importance of accurate approximations and robust noise distribution modeling.
Denise Tampieri, Giovanni De Toni, Paolo Giudici
The University of Edinburgh, Edinburgh, UK · ⋆Work conducted during the author’s MSc at the University of Trento, Trento, Italy. · Fondazione Bruno Kessler, Trento, Italy +1