People
María Martínez García
Saarland Informatics Campus
Building E1.1, Room 2.26
About me
I am a Postdoctoral Researcher in Prof. Isabel Valera’s group in Saarland University since January 2025. I hold a degree in Telecommunications Engineering from Vigo University and a Master’s in Machine Learning applied to Health from University Carlos III in Madrid (UC3M). In 2024, I completed my Ph.D. in Probabilistic Machine Learning applied to Personalized Medicine and Genetics at UC3M, under the supervision of Assoc. Prof. Pablo M. Olmos. During my Ph.D., I also worked as a predoctoral researcher at the Instituto de Investigación Sanitaria Gregorio Marañón (IiSGM) with Dr. Carolina Martínez Laperche, supported by the Intramural Grant.
My research focuses on developing probabilistic machine learning methods for representation learning and dimensionality reduction, particularly in analyzing complex, high-dimensional data with continuous and discrete Variational Autoencoders. Throughout my Ph.D., I combined theoretical and applied research, aiming to bridge the gap between machine learning advancements and real-world clinical applications. This approach allowed me to tackle medical challenges while contributing to machine learning methodology. Now, my focus is on further improving these methods to develop models that are flexible, interpretable and robust.
Publications
2026
Majumdar, Ayan; Kanubala, Deborah Dormah; Gupta, Kavya; Valera, Isabel
A Causal Framework to Measure and Mitigate Non-binary Treatment Discrimination Journal Article
In: CoRR, vol. abs/2503.22454, 2026.
Abstract | Links | BibTeX | Tags: ayanm, deborah, isabel, kavya, saml
@article{DBLP:journals/corr/abs-2503-22454,
title = {A Causal Framework to Measure and Mitigate Non-binary Treatment Discrimination},
author = {Ayan Majumdar and Deborah Dormah Kanubala and Kavya Gupta and Isabel Valera},
url = {https://doi.org/10.48550/arXiv.2503.22454},
doi = {10.48550/ARXIV.2503.22454},
year = {2026},
date = {2026-03-19},
urldate = {2026-03-19},
journal = {CoRR},
volume = {abs/2503.22454},
abstract = {Fairness studies of algorithmic decision-making systems often simplify complex decision processes, such as bail or lending decisions, into binary classification tasks (e.g., approve or not approve). However, these approaches overlook that such decisions are not inherently binary; they also involve non-binary treatment decisions (e.g., loan or bail terms) that can influence the downstream outcomes (e.g., loan repayment or reoffending). We argue that treatment decisions are integral to the decision-making process and, therefore, should be central to fairness analyses. Consequently, we propose a causal framework that extends and complements existing fairness notions by explicitly distinguishing between decision-subjects’ covariates and the treatment decisions. Our framework leverages path-specific counterfactual reasoning to: (i) measure treatment disparity and its downstream effects in historical data; and (ii) mitigate the impact of past unfair treatment decisions when automating decision-making. We use our framework to empirically analyze four widely used loan approval datasets to reveal potential disparity in non-binary treatment decisions and their discriminatory impact on outcomes, highlighting the need to incorporate treatment decisions in fairness assessments. Finally, by intervening in treatment decisions, we show that our framework effectively mitigates treatment discrimination from historical loan approval data to ensure fair risk score estimation and (non-binary) decision-making processes that benefit all stakeholders.},
keywords = {ayanm, deborah, isabel, kavya, saml},
pubstate = {published},
tppubtype = {article}
}
Fairness studies of algorithmic decision-making systems often simplify complex decision processes, such as bail or lending decisions, into binary classification tasks (e.g., approve or not approve). However, these approaches overlook that such decisions are not inherently binary; they also involve non-binary treatment decisions (e.g., loan or bail terms) that can influence the downstream outcomes (e.g., loan repayment or reoffending). We argue that treatment decisions are integral to the decision-making process and, therefore, should be central to fairness analyses. Consequently, we propose a causal framework that extends and complements existing fairness notions by explicitly distinguishing between decision-subjects’ covariates and the treatment decisions. Our framework leverages path-specific counterfactual reasoning to: (i) measure treatment disparity and its downstream effects in historical data; and (ii) mitigate the impact of past unfair treatment decisions when automating decision-making. We use our framework to empirically analyze four widely used loan approval datasets to reveal potential disparity in non-binary treatment decisions and their discriminatory impact on outcomes, highlighting the need to incorporate treatment decisions in fairness assessments. Finally, by intervening in treatment decisions, we show that our framework effectively mitigates treatment discrimination from historical loan approval data to ensure fair risk score estimation and (non-binary) decision-making processes that benefit all stakeholders.
Vo, Huyen Thuc Khanh; Valera, Isabel
Hellinger Multimodal Variational Autoencoders Proceedings Article Spotlight
In: The 29th International Conference on Artificial Intelligence and Statistics, 2026.
Abstract | Links | BibTeX | Tags: huyen, isabel, saml, spotlight
@inproceedings{<LineBreak>vo2026hellinger,
title = {Hellinger Multimodal Variational Autoencoders},
author = {Huyen Thuc Khanh Vo and Isabel Valera},
url = {https://openreview.net/forum?id=mxHyYltMUa},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
booktitle = {The 29th International Conference on Artificial Intelligence and Statistics},
abstract = {Multimodal variational autoencoders (VAEs) are widely used for weakly supervised generative learning with multiple modalities. Predominant methods aggregate unimodal inference distributions using either a product of experts (PoE), a mixture of experts (MoE), or their combinations to approximate the joint posterior. In this work, we revisit multimodal inference through the lens of probabilistic opinion pooling, an optimization-based approach. We start from Hölder pooling with α=0.5, which corresponds to the unique symmetric member of the α-divergence family, and derive a moment-matching approximation, termed Hellinger. We then leverage such an approximation to propose HELVAE, a multimodal VAE that avoids sub-sampling, yielding an efficient yet effective model that: (i) learns more expressive latent representations as additional modalities are observed; and (ii) empirically achieves better trade-offs between generative coherence and quality, outperforming state-of-the-art multimodal VAE models.},
keywords = {huyen, isabel, saml, spotlight},
pubstate = {published},
tppubtype = {inproceedings}
}
Multimodal variational autoencoders (VAEs) are widely used for weakly supervised generative learning with multiple modalities. Predominant methods aggregate unimodal inference distributions using either a product of experts (PoE), a mixture of experts (MoE), or their combinations to approximate the joint posterior. In this work, we revisit multimodal inference through the lens of probabilistic opinion pooling, an optimization-based approach. We start from Hölder pooling with α=0.5, which corresponds to the unique symmetric member of the α-divergence family, and derive a moment-matching approximation, termed Hellinger. We then leverage such an approximation to propose HELVAE, a multimodal VAE that avoids sub-sampling, yielding an efficient yet effective model that: (i) learns more expressive latent representations as additional modalities are observed; and (ii) empirically achieves better trade-offs between generative coherence and quality, outperforming state-of-the-art multimodal VAE models.
Gupta, Kavya; Kalampalikis, Nektarios; Heitz, Christoph; Valera, Isabel
First-See-Then-Design: A Multi-Stakeholder View for Optimal Performance-Fairness Trade-Offs Journal Article
In: arXiv preprint arXiv:2604.14035, 2026.
Abstract | Links | BibTeX | Tags: isabel, nektarios, saml
@article{gupta2026first,
title = {First-See-Then-Design: A Multi-Stakeholder View for Optimal Performance-Fairness Trade-Offs},
author = {Kavya Gupta and Nektarios Kalampalikis and Christoph Heitz and Isabel Valera},
url = {https://arxiv.org/abs/2604.14035},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
journal = {arXiv preprint arXiv:2604.14035},
abstract = {Fairness in algorithmic decision-making is often defined in the predictive space, where predictive performance - used as a proxy for decision-maker (DM) utility - is traded off against prediction-based fairness notions, such as demographic parity or equality of opportunity. This perspective, however, ignores how predictions translate into decisions and ultimately into utilities and welfare for both DM and decision subjects (DS), as well as their allocation across social-salient groups.
In this paper, we propose a multi-stakeholder framework for fair algorithmic decision-making grounded in welfare economics and distributive justice, explicitly modeling the utilities of both the DM and DS, and defining fairness via a social planner's utility that captures inequalities in DS utilities across groups under different justice-based fairness notions (e.g., Egalitarian, Rawlsian). We formulate fair decision-making as a post-hoc multi-objective optimization problem, characterizing the achievable performance-fairness trade-offs in the two-dimensional utility space of DM utility and the social planner's utility, under different decision policy classes (deterministic vs. stochastic, shared vs. group-specific). Using the proposed framework, we then identify conditions (in terms of the stakeholders' utilities) under which stochastic policies are more optimal than deterministic ones, and empirically demonstrate that simple stochastic policies can yield superior performance-fairness trade-offs by leveraging outcome uncertainty. Overall, we advocate a shift from prediction-centric fairness to a transparent, justice-based, multi-stakeholder approach that supports the collaborative design of decision-making policies.},
keywords = {isabel, nektarios, saml},
pubstate = {published},
tppubtype = {article}
}
Fairness in algorithmic decision-making is often defined in the predictive space, where predictive performance - used as a proxy for decision-maker (DM) utility - is traded off against prediction-based fairness notions, such as demographic parity or equality of opportunity. This perspective, however, ignores how predictions translate into decisions and ultimately into utilities and welfare for both DM and decision subjects (DS), as well as their allocation across social-salient groups.
In this paper, we propose a multi-stakeholder framework for fair algorithmic decision-making grounded in welfare economics and distributive justice, explicitly modeling the utilities of both the DM and DS, and defining fairness via a social planner's utility that captures inequalities in DS utilities across groups under different justice-based fairness notions (e.g., Egalitarian, Rawlsian). We formulate fair decision-making as a post-hoc multi-objective optimization problem, characterizing the achievable performance-fairness trade-offs in the two-dimensional utility space of DM utility and the social planner's utility, under different decision policy classes (deterministic vs. stochastic, shared vs. group-specific). Using the proposed framework, we then identify conditions (in terms of the stakeholders' utilities) under which stochastic policies are more optimal than deterministic ones, and empirically demonstrate that simple stochastic policies can yield superior performance-fairness trade-offs by leveraging outcome uncertainty. Overall, we advocate a shift from prediction-centric fairness to a transparent, justice-based, multi-stakeholder approach that supports the collaborative design of decision-making policies.
In this paper, we propose a multi-stakeholder framework for fair algorithmic decision-making grounded in welfare economics and distributive justice, explicitly modeling the utilities of both the DM and DS, and defining fairness via a social planner's utility that captures inequalities in DS utilities across groups under different justice-based fairness notions (e.g., Egalitarian, Rawlsian). We formulate fair decision-making as a post-hoc multi-objective optimization problem, characterizing the achievable performance-fairness trade-offs in the two-dimensional utility space of DM utility and the social planner's utility, under different decision policy classes (deterministic vs. stochastic, shared vs. group-specific). Using the proposed framework, we then identify conditions (in terms of the stakeholders' utilities) under which stochastic policies are more optimal than deterministic ones, and empirically demonstrate that simple stochastic policies can yield superior performance-fairness trade-offs by leveraging outcome uncertainty. Overall, we advocate a shift from prediction-centric fairness to a transparent, justice-based, multi-stakeholder approach that supports the collaborative design of decision-making policies.
Valdrighi, Giovani; Valera, Isabel; Raimundo, Marcos Medeiros
Long-term Fairness with Selective Labels Miscellaneous
2026.
Abstract | Links | BibTeX | Tags: isabel, saml
@misc{valdrighi2026longtermfairnessselectivelabels,
title = {Long-term Fairness with Selective Labels},
author = {Giovani Valdrighi and Isabel Valera and Marcos Medeiros Raimundo},
url = {https://arxiv.org/abs/2605.22291},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
abstract = {Long-term fairness algorithms aim to satisfy fairness beyond static and short-term notions by accounting for the dynamics between decision-making policies and population behavior. Most previous approaches evaluate performance and fairness measures from observable features and a label, which is assumed to be fully observed. However, in scenarios such as hiring or lending, the labels (e.g., ability to repay the loan) are selective labels as they are only revealed based on positive decisions (e.g., when a loan is granted). In this paper, we study long-term fairness in the selective labels setting and analytically show that naive solutions do not guarantee fairness. To address this gap, we then introduce a novel framework that leverages both the observed data and a label predictor model to estimate the true fairness measure value by decomposing it into the observed fairness and bias from label predictions. This allows us to derive sufficient conditions to satisfy true fairness from observable quantities by using the confidence in the predictor model. Finally, we rely on our theoretical results to propose a novel reinforcement learning algorithm for effective long-term fair decision-making with selective labels. In semisynthetic environments, the proposed algorithm reached comparable fairness and performance to an agent with oracle access to the true labels.},
keywords = {isabel, saml},
pubstate = {published},
tppubtype = {misc}
}
Long-term fairness algorithms aim to satisfy fairness beyond static and short-term notions by accounting for the dynamics between decision-making policies and population behavior. Most previous approaches evaluate performance and fairness measures from observable features and a label, which is assumed to be fully observed. However, in scenarios such as hiring or lending, the labels (e.g., ability to repay the loan) are selective labels as they are only revealed based on positive decisions (e.g., when a loan is granted). In this paper, we study long-term fairness in the selective labels setting and analytically show that naive solutions do not guarantee fairness. To address this gap, we then introduce a novel framework that leverages both the observed data and a label predictor model to estimate the true fairness measure value by decomposing it into the observed fairness and bias from label predictions. This allows us to derive sufficient conditions to satisfy true fairness from observable quantities by using the confidence in the predictor model. Finally, we rely on our theoretical results to propose a novel reinforcement learning algorithm for effective long-term fair decision-making with selective labels. In semisynthetic environments, the proposed algorithm reached comparable fairness and performance to an agent with oracle access to the true labels.
Budde, Lena Marie; Majumdar, Ayan; Uth, Richard; Langer, Markus; Valera, Isabel
From Universal to Individualized Actionability: Revisiting Personalization in Algorithmic Recourse Miscellaneous
2026.
Abstract | Links | BibTeX | Tags: isabel, saml
@misc{budde2026universalindividualizedactionabilityrevisiting,
title = {From Universal to Individualized Actionability: Revisiting Personalization in Algorithmic Recourse},
author = {Lena Marie Budde and Ayan Majumdar and Richard Uth and Markus Langer and Isabel Valera},
url = {https://arxiv.org/abs/2604.08030},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
abstract = {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.},
keywords = {isabel, saml},
pubstate = {published},
tppubtype = {misc}
}
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.
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.
Martínez-García, María; Alvarez, Ricardo Vazquez; Lancho, Alejandro; Olmos, Pablo M.; Valera, Isabel
A Probabilistic Hard Concept Bottleneck for Steerable Generative Models Proceedings Article
In: The Fourteenth International Conference on Learning Representations, 2026.
Abstract | Links | BibTeX | Tags: isabel, maria, pablo, saml
@inproceedings{<LineBreak>martinez-garcia2026a,
title = {A Probabilistic Hard Concept Bottleneck for Steerable Generative Models},
author = {María Martínez-García and Ricardo Vazquez Alvarez and Alejandro Lancho and Pablo M. Olmos and Isabel Valera},
url = {https://openreview.net/forum?id=Kcb6WufAco},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
booktitle = {The Fourteenth International Conference on Learning Representations},
abstract = {Concept Bottleneck Generative Models (CBGMs) incorporate a human-interpretable concept bottleneck layer, which makes them interpretable and steerable. However, designing such a layer for generative models poses the same challenges as for concept bottleneck models in a supervised context, if not greater ones. Deterministic mappings from the model inner representations to soft concepts in existing CBGMs: (i) limit steerable generation to modifying concepts in existing inputs; and, more importantly, (ii) are susceptible to concept leakage, which hinders their steerability. To address these limitations, we first introduce the Variational Hard Concept Bottleneck (VHCB) layer. The VHCB maps probabilistic estimates of binary latent variables to hard concepts, which have been shown to mitigate leakage. Remarkably, its probabilistic formulation enables direct generation from a specified set of concepts. Second, we propose a systematic evaluation framework for assessing the steerability of CBGMs across various tasks (e.g., activating and deactivating concepts). Our framework which allows us to empirically demonstrate that the VHCB layer consistently improves steerability.},
keywords = {isabel, maria, pablo, saml},
pubstate = {published},
tppubtype = {inproceedings}
}
Concept Bottleneck Generative Models (CBGMs) incorporate a human-interpretable concept bottleneck layer, which makes them interpretable and steerable. However, designing such a layer for generative models poses the same challenges as for concept bottleneck models in a supervised context, if not greater ones. Deterministic mappings from the model inner representations to soft concepts in existing CBGMs: (i) limit steerable generation to modifying concepts in existing inputs; and, more importantly, (ii) are susceptible to concept leakage, which hinders their steerability. To address these limitations, we first introduce the Variational Hard Concept Bottleneck (VHCB) layer. The VHCB maps probabilistic estimates of binary latent variables to hard concepts, which have been shown to mitigate leakage. Remarkably, its probabilistic formulation enables direct generation from a specified set of concepts. Second, we propose a systematic evaluation framework for assessing the steerability of CBGMs across various tasks (e.g., activating and deactivating concepts). Our framework which allows us to empirically demonstrate that the VHCB layer consistently improves steerability.
Vo, Huyen; Martínez-García, María; Valera, Isabel
Hölder++: Improving the Quality-Coherence Trade-off in Multimodal VAEs Miscellaneous
2026.
Abstract | Links | BibTeX | Tags: huyen, isabel, maria, saml
@misc{vo2026holderimprovingqualitycoherencetradeoff,
title = {Hölder++: Improving the Quality-Coherence Trade-off in Multimodal VAEs},
author = {Huyen Vo and María Martínez-García and Isabel Valera},
url = {https://arxiv.org/abs/2606.13381},
year = {2026},
date = {2026-01-01},
urldate = {2026-01-01},
abstract = {Existing approaches for multimodal variational autoencoders (VAEs) face a trade-off between generative quality and coherence-i.e., they struggle to generate realistic and diverse samples that, at the same time, are semantically consistent across modalities. A recent work shows that using a simple approximation to Hölder pooling as an aggregation method improves coherence over the SOTA MMVAE+, despite assuming a single shared representation across all modalities. Yet, it slightly compromises sample diversity. Inspired by this insight, we propose Hölder++, a novel multimodal VAE that improves the generative quality-coherence trade-off through: (i) the first implementation of Hölder pooling without any approximation for multimodal VAEs; (ii) an extended architecture that models distinct shared and private (i.e., modality-specific) representations (Hölder+); and (iii) hierarchical inference that further enhances the disentanglement between the shared and private representations (Hölder++). Our experiments corroborate that Hölder++ consistently improves the generative quality-coherence trade-off, yields more structured latent spaces, and learns shared representations that are informative for downstream tasks.},
keywords = {huyen, isabel, maria, saml},
pubstate = {published},
tppubtype = {misc}
}
Existing approaches for multimodal variational autoencoders (VAEs) face a trade-off between generative quality and coherence-i.e., they struggle to generate realistic and diverse samples that, at the same time, are semantically consistent across modalities. A recent work shows that using a simple approximation to Hölder pooling as an aggregation method improves coherence over the SOTA MMVAE+, despite assuming a single shared representation across all modalities. Yet, it slightly compromises sample diversity. Inspired by this insight, we propose Hölder++, a novel multimodal VAE that improves the generative quality-coherence trade-off through: (i) the first implementation of Hölder pooling without any approximation for multimodal VAEs; (ii) an extended architecture that models distinct shared and private (i.e., modality-specific) representations (Hölder+); and (iii) hierarchical inference that further enhances the disentanglement between the shared and private representations (Hölder++). Our experiments corroborate that Hölder++ consistently improves the generative quality-coherence trade-off, yields more structured latent spaces, and learns shared representations that are informative for downstream tasks.
2025
Azime, Israel Abebe; Kanubala, Deborah D.; Afonja, Tejumade; Fritz, Mario; Valera, Isabel; Klakow, Dietrich; Slusallek, Philipp
Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches Bachelor Thesis
2025.
Abstract | Links | BibTeX | Tags: isabel, saml
@bachelorthesis{nokey,
title = {Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches},
author = {Israel Abebe Azime and Deborah D. Kanubala and Tejumade Afonja and Mario Fritz and Isabel Valera and Dietrich Klakow and Philipp Slusallek},
url = {https://arxiv.org/pdf/2508.21512},
year = {2025},
date = {2025-08-29},
urldate = {2025-08-29},
abstract = {Large Language Models (LLMs) are increasingly employed in high-stakes decision-making tasks, such as loan approvals. While their applications expand across domains, LLMs struggle to process tabular data, ensuring fairness and delivering reliable predictions. In this work, we assess the performance and fairness of LLMs on serialized loan approval datasets from three geographically distinct regions: Ghana, Germany, and the United States. Our evaluation focuses on the model’s zero-shot and in-context learning (ICL) capabilities. Our results reveal that the choice of serialization1 format significantly affects both performance and fairness in LLMs, with certain formats such as GReaT and LIFT yielding higher F1 scores but exacerbating fairness disparities. Notably, while ICL improved model performance by 4.9-59.6% relative to zero-shot baselines, its effect on fairness varied considerably across datasets. Our work underscores the importance of effective tabular data representation methods and fairness-aware models to improve the reliability of LLMs in financial decision-making},
keywords = {isabel, saml},
pubstate = {published},
tppubtype = {bachelorthesis}
}
Large Language Models (LLMs) are increasingly employed in high-stakes decision-making tasks, such as loan approvals. While their applications expand across domains, LLMs struggle to process tabular data, ensuring fairness and delivering reliable predictions. In this work, we assess the performance and fairness of LLMs on serialized loan approval datasets from three geographically distinct regions: Ghana, Germany, and the United States. Our evaluation focuses on the model’s zero-shot and in-context learning (ICL) capabilities. Our results reveal that the choice of serialization1 format significantly affects both performance and fairness in LLMs, with certain formats such as GReaT and LIFT yielding higher F1 scores but exacerbating fairness disparities. Notably, while ICL improved model performance by 4.9-59.6% relative to zero-shot baselines, its effect on fairness varied considerably across datasets. Our work underscores the importance of effective tabular data representation methods and fairness-aware models to improve the reliability of LLMs in financial decision-making
Javaloy, Adrián; Vergari, Antonio; Valera, Isabel
COPA: Comparing the Incomparable to Explore the Pareto Front Journal Article
In: CoRR, vol. abs/2503.14321, 2025.
Abstract | Links | BibTeX | Tags: adrian, isabel, saml
@article{DBLP:journals/corr/abs-2503-14321,
title = {COPA: Comparing the Incomparable to Explore the Pareto Front},
author = {Adrián Javaloy and Antonio Vergari and Isabel Valera},
url = {https://doi.org/10.48550/arXiv.2503.14321},
doi = {10.48550/ARXIV.2503.14321},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {CoRR},
volume = {abs/2503.14321},
abstract = {In machine learning (ML), we often need to choose one among hundreds of trained ML models at hand, based on various objectives such as accuracy, robustness, fairness or scalability. However, it is often unclear how to compare, aggregate and, ultimately, trade-off these objectives, making it a time-consuming task that requires expert knowledge, as objectives may be measured in different units and scales. In this work, we investigate how objectives can be automatically normalized and aggregated to systematically help the user navigate their Pareto front. To this end, we make incomparable objectives comparable using their cumulative functions, approximated by their relative rankings. As a result, our proposed approach, COPA, can aggregate them while matching user-specific preferences, allowing practitioners to meaningfully navigate and search for models in the Pareto front. We demonstrate the potential impact of COPA in both model selection and benchmarking tasks across diverse ML areas such as fair ML, domain generalization, AutoML and foundation models, where classical ways to normalize and aggregate objectives fall short.},
keywords = {adrian, isabel, saml},
pubstate = {published},
tppubtype = {article}
}
In machine learning (ML), we often need to choose one among hundreds of trained ML models at hand, based on various objectives such as accuracy, robustness, fairness or scalability. However, it is often unclear how to compare, aggregate and, ultimately, trade-off these objectives, making it a time-consuming task that requires expert knowledge, as objectives may be measured in different units and scales. In this work, we investigate how objectives can be automatically normalized and aggregated to systematically help the user navigate their Pareto front. To this end, we make incomparable objectives comparable using their cumulative functions, approximated by their relative rankings. As a result, our proposed approach, COPA, can aggregate them while matching user-specific preferences, allowing practitioners to meaningfully navigate and search for models in the Pareto front. We demonstrate the potential impact of COPA in both model selection and benchmarking tasks across diverse ML areas such as fair ML, domain generalization, AutoML and foundation models, where classical ways to normalize and aggregate objectives fall short.
Almodóvar, Alejandro; Javaloy, Adrián; Parras, Juan; Zazo, Santiago; Valera, Isabel
DeCaFlow: A Deconfounding Causal Generative Model Journal Article Spotlight
In: CoRR, vol. abs/2503.15114, 2025.
Abstract | Links | BibTeX | Tags: adrian, isabel, saml, spotlight
@article{DBLP:journals/corr/abs-2503-15114,
title = {DeCaFlow: A Deconfounding Causal Generative Model},
author = {Alejandro Almodóvar and Adrián Javaloy and Juan Parras and Santiago Zazo and Isabel Valera},
url = {https://doi.org/10.48550/arXiv.2503.15114},
doi = {10.48550/ARXIV.2503.15114},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {CoRR},
volume = {abs/2503.15114},
abstract = {We introduce DeCaFlow, a deconfounding causal generative model. Training once per dataset using just observational data and the underlying causal graph, DeCaFlow enables accurate causal inference on continuous variables under the presence of hidden confounders. Specifically, we extend previous results on causal estimation under hidden confounding to show that a single instance of DeCaFlow provides correct estimates for all causal queries identifiable with do-calculus, leveraging proxy variables to adjust for the causal effects when do-calculus alone is insufficient. Moreover, we show that counterfactual queries are identifiable as long as their interventional counterparts are identifiable, and thus are also correctly estimated by DeCaFlow. Our empirical results on diverse settings (including the Ecoli70 dataset, with 3 independent hidden confounders, tens of observed variables and hundreds of causal queries) show that DeCaFlow outperforms existing approaches, while demonstrating its out-of-the-box applicability to any given causal graph},
keywords = {adrian, isabel, saml, spotlight},
pubstate = {published},
tppubtype = {article}
}
We introduce DeCaFlow, a deconfounding causal generative model. Training once per dataset using just observational data and the underlying causal graph, DeCaFlow enables accurate causal inference on continuous variables under the presence of hidden confounders. Specifically, we extend previous results on causal estimation under hidden confounding to show that a single instance of DeCaFlow provides correct estimates for all causal queries identifiable with do-calculus, leveraging proxy variables to adjust for the causal effects when do-calculus alone is insufficient. Moreover, we show that counterfactual queries are identifiable as long as their interventional counterparts are identifiable, and thus are also correctly estimated by DeCaFlow. Our empirical results on diverse settings (including the Ecoli70 dataset, with 3 independent hidden confounders, tens of observed variables and hundreds of causal queries) show that DeCaFlow outperforms existing approaches, while demonstrating its out-of-the-box applicability to any given causal graph
Cinquini, Martina; Beretta, Isacco; Ruggieri, Salvatore; Valera, Isabel
A Practical Approach to Causal Inference over Time Proceedings Article
In: Walsh, Toby; Shah, Julie; Kolter, Zico (Ed.): AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA, pp. 14832–14839, AAAI Press, 2025.
Abstract | Links | BibTeX | Tags: isabel, saml
@inproceedings{DBLP:conf/aaai/CinquiniBRV25,
title = {A Practical Approach to Causal Inference over Time},
author = {Martina Cinquini and Isacco Beretta and Salvatore Ruggieri and Isabel Valera},
editor = {Toby Walsh and Julie Shah and Zico Kolter},
url = {https://doi.org/10.1609/aaai.v39i14.33626},
doi = {10.1609/AAAI.V39I14.33626},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
booktitle = {AAAI-25, Sponsored by the Association for the Advancement of Artificial
Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA},
pages = {14832–14839},
publisher = {AAAI Press},
abstract = {In this paper, we focus on estimating the causal effect of an intervention over time on a dynamical system. To that end, we formally define causal interventions and their effects over time on discrete-time stochastic processes (DSPs). Then, we show under which conditions the equilibrium states of a DSP, both before and after a causal intervention, can be captured by a structural causal model (SCM). With such an equivalence at hand, we provide an explicit mapping from vector autoregressive models (VARs), broadly applied in econometrics, to linear, but potentially cyclic and/or affected by unmeasured confounders, SCMs. The resulting causal VAR framework allows us to perform causal inference over time from observational time series data. Our experiments on synthetic and real-world datasets show that the proposed framework achieves strong performance in terms of observational forecasting while enabling accurate estimation of the causal effect of interventions on dynamical systems. We demonstrate, through a case study, the potential practical questions that can be addressed using the proposed causal VAR framework.},
keywords = {isabel, saml},
pubstate = {published},
tppubtype = {inproceedings}
}
In this paper, we focus on estimating the causal effect of an intervention over time on a dynamical system. To that end, we formally define causal interventions and their effects over time on discrete-time stochastic processes (DSPs). Then, we show under which conditions the equilibrium states of a DSP, both before and after a causal intervention, can be captured by a structural causal model (SCM). With such an equivalence at hand, we provide an explicit mapping from vector autoregressive models (VARs), broadly applied in econometrics, to linear, but potentially cyclic and/or affected by unmeasured confounders, SCMs. The resulting causal VAR framework allows us to perform causal inference over time from observational time series data. Our experiments on synthetic and real-world datasets show that the proposed framework achieves strong performance in terms of observational forecasting while enabling accurate estimation of the causal effect of interventions on dynamical systems. We demonstrate, through a case study, the potential practical questions that can be addressed using the proposed causal VAR framework.
Peis, Ignacio; Koyuncu, Batuhan; Valera, Isabel; Frellsen, Jes
Hyper-Transforming Latent Diffusion Models Journal Article
In: CoRR, vol. abs/2504.16580, 2025.
Abstract | Links | BibTeX | Tags: batu, isabel, saml
@article{DBLP:journals/corr/abs-2504-16580,
title = {Hyper-Transforming Latent Diffusion Models},
author = {Ignacio Peis and Batuhan Koyuncu and Isabel Valera and Jes Frellsen},
url = {https://doi.org/10.48550/arXiv.2504.16580},
doi = {10.48550/ARXIV.2504.16580},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {CoRR},
volume = {abs/2504.16580},
abstract = {We introduce a novel generative framework for functions by integrating Implicit Neural Representations (INRs) and Transformer-based hypernetworks into latent variable models. Unlike prior approaches that rely on MLP-based hypernetworks with scalability limitations, our method employs a Transformer-based decoder to generate INR parameters from latent variables, addressing both representation capacity and computational efficiency. Our framework extends latent diffusion models (LDMs) to INR generation by replacing standard decoders with a Transformer-based hypernetwork, which can be trained either from scratch or via hyper-transforming: a strategy that fine-tunes only the decoder while freezing the pre-trained latent space. This enables efficient adaptation of existing generative models to INR-based representations without requiring full retraining. We validate our approach across multiple modalities, demonstrating improved scalability, expressiveness, and generalization over existing INR-based generative models. Our findings establish a unified and flexible framework for learning structured function representations.},
keywords = {batu, isabel, saml},
pubstate = {published},
tppubtype = {article}
}
We introduce a novel generative framework for functions by integrating Implicit Neural Representations (INRs) and Transformer-based hypernetworks into latent variable models. Unlike prior approaches that rely on MLP-based hypernetworks with scalability limitations, our method employs a Transformer-based decoder to generate INR parameters from latent variables, addressing both representation capacity and computational efficiency. Our framework extends latent diffusion models (LDMs) to INR generation by replacing standard decoders with a Transformer-based hypernetwork, which can be trained either from scratch or via hyper-transforming: a strategy that fine-tunes only the decoder while freezing the pre-trained latent space. This enables efficient adaptation of existing generative models to INR-based representations without requiring full retraining. We validate our approach across multiple modalities, demonstrating improved scalability, expressiveness, and generalization over existing INR-based generative models. Our findings establish a unified and flexible framework for learning structured function representations.
Kanubala, Deborah Dormah; Valera, Isabel
On the Misalignment Between Legal Notions and Statistical Metrics of Intersectional Fairness Journal Article
In: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 2025.
Abstract | Links | BibTeX | Tags: deborah, isabel, saml
@article{Kanubala2025OnTM,
title = {On the Misalignment Between Legal Notions and Statistical Metrics of Intersectional Fairness},
author = {Deborah Dormah Kanubala and Isabel Valera},
url = {https://api.semanticscholar.org/CorpusID:282175621},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society},
abstract = {Intersectional (un)fairness, as conceptualized in legal and social theory, emphasizes the non-additive and structurally complex nature of discrimination against individuals at the intersection of multiple sensitive attributes (such as race, gender, etc). Recent works have proposed statistical metrics for intersectional fairness by estimating disparities across groups of individuals sharing two or more sensitive attributes. However, it is unclear if these metrics detect uniquely intersectional discrimination. We therefore pose the following question, Do current statistical intersectional metrics detect the non-additive discrimination highlighted by intersectionality theory? More specifically, to answer this, we run controlled synthetic data experiments that explicitly allow us to control for single, multiple, intersectional, and compounded forms of discrimination. Our analyses show that current statistical metrics for intersectional fairness behave more like multi-attribute disparity measures. Specifically, they respond more strongly to additive or compounded biases than to non-additive interaction effects. While they effectively capture disparities across multiple sensitive attributes, they often fail to detect uniquely intersectional discrimination. These findings reveal a fundamental misalignment between existing intersectional fairness metrics and the legal and theoretical foundations of intersectionality. We argue that if intersectional fairness metrics are to be deemed truly intersectional, they must be explicitly designed to account for the structural, non-additive nature of intersectional discrimination.},
keywords = {deborah, isabel, saml},
pubstate = {published},
tppubtype = {article}
}
Intersectional (un)fairness, as conceptualized in legal and social theory, emphasizes the non-additive and structurally complex nature of discrimination against individuals at the intersection of multiple sensitive attributes (such as race, gender, etc). Recent works have proposed statistical metrics for intersectional fairness by estimating disparities across groups of individuals sharing two or more sensitive attributes. However, it is unclear if these metrics detect uniquely intersectional discrimination. We therefore pose the following question, Do current statistical intersectional metrics detect the non-additive discrimination highlighted by intersectionality theory? More specifically, to answer this, we run controlled synthetic data experiments that explicitly allow us to control for single, multiple, intersectional, and compounded forms of discrimination. Our analyses show that current statistical metrics for intersectional fairness behave more like multi-attribute disparity measures. Specifically, they respond more strongly to additive or compounded biases than to non-additive interaction effects. While they effectively capture disparities across multiple sensitive attributes, they often fail to detect uniquely intersectional discrimination. These findings reveal a fundamental misalignment between existing intersectional fairness metrics and the legal and theoretical foundations of intersectionality. We argue that if intersectional fairness metrics are to be deemed truly intersectional, they must be explicitly designed to account for the structural, non-additive nature of intersectional discrimination.
Kalampalikis, Nektarios; Gupta, Kavya; Vitanov, Georgi; Valera, Isabel
Towards Reasonable Concept Bottleneck Models Journal Article
In: CoRR, vol. abs/2506.05014, 2025.
Abstract | Links | BibTeX | Tags: isabel, kavya, nektarios, saml
@article{DBLP:journals/corr/abs-2506-05014,
title = {Towards Reasonable Concept Bottleneck Models},
author = {Nektarios Kalampalikis and Kavya Gupta and Georgi Vitanov and Isabel Valera},
url = {https://doi.org/10.48550/arXiv.2506.05014},
doi = {10.48550/ARXIV.2506.05014},
year = {2025},
date = {2025-01-01},
urldate = {2025-01-01},
journal = {CoRR},
volume = {abs/2506.05014},
abstract = {In this paper, we propose textbf{C}oncept textbf{REA}soning textbf{M}odels (CREAM), a novel family of Concept Bottleneck Models (CBMs) that: (i) explicitly encodes concept-concept ({texttt{C-C}}) and concept-task ({texttt{C$rightarrow$Y}}) relationships to enforce a desired model reasoning; and (ii) use a regularized side-channel to achieve competitive task performance, while keeping high concept importance. Specifically, CREAM architecturally embeds (bi)directed concept-concept, and concept to task relationships specified by a human expert, while severing undesired information flows (e.g., to handle mutually exclusive concepts). Moreover, CREAM integrates a black-box side-channel that is regularized to encourage task predictions to be grounded in the relevant concepts, thereby utilizing the side-channel only when necessary to enhance performance. Our experiments show that: (i) CREAM mainly relies on concepts while achieving task performance on par with black-box models; and (ii) the embedded {texttt{C-C}} and {texttt{C$rightarrow$Y}} relationships ease model interventions and mitigate concept leakage.},
keywords = {isabel, kavya, nektarios, saml},
pubstate = {published},
tppubtype = {article}
}
In this paper, we propose textbf{C}oncept textbf{REA}soning textbf{M}odels (CREAM), a novel family of Concept Bottleneck Models (CBMs) that: (i) explicitly encodes concept-concept ({texttt{C-C}}) and concept-task ({texttt{C$rightarrow$Y}}) relationships to enforce a desired model reasoning; and (ii) use a regularized side-channel to achieve competitive task performance, while keeping high concept importance. Specifically, CREAM architecturally embeds (bi)directed concept-concept, and concept to task relationships specified by a human expert, while severing undesired information flows (e.g., to handle mutually exclusive concepts). Moreover, CREAM integrates a black-box side-channel that is regularized to encourage task predictions to be grounded in the relevant concepts, thereby utilizing the side-channel only when necessary to enhance performance. Our experiments show that: (i) CREAM mainly relies on concepts while achieving task performance on par with black-box models; and (ii) the embedded {texttt{C-C}} and {texttt{C$rightarrow$Y}} relationships ease model interventions and mitigate concept leakage.
2024
Kanubala, Deborah Dormah; Valera, Isabel; Gupta, Kavya
Fairness Beyond Binary Decisions: a Case Study on German Credit Proceedings Article
In: Cerrato, Mattia; Coronel, Alesia Vallenas; Ahrweiler, Petra; Loi, Michele; Pechenizkiy, Mykola; Tamò-Larrieux, Aurelia (Ed.): Proceedings of the 3rd European Workshop on Algorithmic Fairness, Mainz, Germany, July 1st to 3rd, 2024, CEUR-WS.org, 2024.
Abstract | Links | BibTeX | Tags: deborah, isabel, kavya, saml
@inproceedings{DBLP:conf/ewaf/KanubalaVG24,
title = {Fairness Beyond Binary Decisions: a Case Study on German Credit},
author = {Deborah Dormah Kanubala and Isabel Valera and Kavya Gupta},
editor = {Mattia Cerrato and Alesia Vallenas Coronel and Petra Ahrweiler and Michele Loi and Mykola Pechenizkiy and Aurelia Tamò-Larrieux},
url = {https://ceur-ws.org/Vol-3908/paper_15.pdf},
year = {2024},
date = {2024-01-01},
urldate = {2024-01-01},
booktitle = {Proceedings of the 3rd European Workshop on Algorithmic Fairness,
Mainz, Germany, July 1st to 3rd, 2024},
volume = {3908},
publisher = {CEUR-WS.org},
series = {CEUR Workshop Proceedings},
abstract = {Data-driven approaches are increasingly used to (partially) automate decision-making in credit scoring
by predicting whether an applicant is “creditworthy or not” based on a set of features about the applicant,
such as age and income, along with what we refer here to as treatment decisions, e.g., loan amount and
duration. Existing data-driven approaches for automating and evaluating the accuracy and fairness of
such credit decisions ignore that treatment decisions (here, loan terms) are part of the decision and
thus may be subject to discrimination. This discrimination can propagate to the final outcome (repaid
or not) of positive decisions (granted loans). In this extended abstract, we rely on causal reasoning
and a broadly studied fair machine-learning dataset, the German credit, to i) show that the current fair
data-driven approach neglects discrimination in treatment decisions (i.e., loan terms) and its downstream
consequences on the decision outcome (i.e., ability to repay); and ii) argue for the need to move beyond
binary decisions in fair data-driven decision-making in consequential settings like credit scoring},
keywords = {deborah, isabel, kavya, saml},
pubstate = {published},
tppubtype = {inproceedings}
}
Data-driven approaches are increasingly used to (partially) automate decision-making in credit scoring
by predicting whether an applicant is “creditworthy or not” based on a set of features about the applicant,
such as age and income, along with what we refer here to as treatment decisions, e.g., loan amount and
duration. Existing data-driven approaches for automating and evaluating the accuracy and fairness of
such credit decisions ignore that treatment decisions (here, loan terms) are part of the decision and
thus may be subject to discrimination. This discrimination can propagate to the final outcome (repaid
or not) of positive decisions (granted loans). In this extended abstract, we rely on causal reasoning
and a broadly studied fair machine-learning dataset, the German credit, to i) show that the current fair
data-driven approach neglects discrimination in treatment decisions (i.e., loan terms) and its downstream
consequences on the decision outcome (i.e., ability to repay); and ii) argue for the need to move beyond
binary decisions in fair data-driven decision-making in consequential settings like credit scoring
by predicting whether an applicant is “creditworthy or not” based on a set of features about the applicant,
such as age and income, along with what we refer here to as treatment decisions, e.g., loan amount and
duration. Existing data-driven approaches for automating and evaluating the accuracy and fairness of
such credit decisions ignore that treatment decisions (here, loan terms) are part of the decision and
thus may be subject to discrimination. This discrimination can propagate to the final outcome (repaid
or not) of positive decisions (granted loans). In this extended abstract, we rely on causal reasoning
and a broadly studied fair machine-learning dataset, the German credit, to i) show that the current fair
data-driven approach neglects discrimination in treatment decisions (i.e., loan terms) and its downstream
consequences on the decision outcome (i.e., ability to repay); and ii) argue for the need to move beyond
binary decisions in fair data-driven decision-making in consequential settings like credit scoring
Rateike, Miriam; Valera, Isabel; Forré, Patrick
Designing Long-term Group Fair Policies in Dynamical Systems Proceedings Article
In: The 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT 2024, Rio de Janeiro, Brazil, June 3-6, 2024, pp. 20–50, ACM, 2024.
Abstract | Links | BibTeX | Tags: isabel, saml
@inproceedings{DBLP:conf/fat/RateikeVF24,
title = {Designing Long-term Group Fair Policies in Dynamical Systems},
author = {Miriam Rateike and Isabel Valera and Patrick Forré},
url = {https://doi.org/10.1145/3630106.3658538},
doi = {10.1145/3630106.3658538},
year = {2024},
date = {2024-01-01},
urldate = {2024-01-01},
booktitle = {The 2024 ACM Conference on Fairness, Accountability, and Transparency,
FAccT 2024, Rio de Janeiro, Brazil, June 3-6, 2024},
pages = {20–50},
publisher = {ACM},
abstract = {Neglecting the effect that decisions have on individuals (and thus, on the underlying data distribution) when designing algorithmic decision-making policies may increase inequalities and unfairness in the long term—even if fairness considerations were taken into account in the policy design process. In this paper, we propose a novel framework for studying long-term group fairness in dynamical systems, in which current decisions may affect an individual’s features in the next step, and thus, future decisions. Specifically, our framework allows us to identify a time-independent policy that converges, if deployed, to the targeted fair stationary state of the system in the long-term, independently of the initial data distribution. We model the system dynamics with a time-homogeneous Markov chain and optimize the policy leveraging the Markov Chain Convergence Theorem to ensure unique convergence. Our framework enables the utilization of historical temporal data to tackle challenges associated with delayed feedback when learning long-term fair policies in practice. Importantly, our framework shows that interventions on the data distribution (e.g., subsidies) can be used to achieve policy learning that is both short- and long-term fair. We provide examples of different targeted fair states of the system, encompassing a range of long-term goals for society and policymakers. In semi-synthetic simulations based on real-world datasets, we show how our approach facilitates identifying effective interventions for long-term fairness.},
keywords = {isabel, saml},
pubstate = {published},
tppubtype = {inproceedings}
}
Neglecting the effect that decisions have on individuals (and thus, on the underlying data distribution) when designing algorithmic decision-making policies may increase inequalities and unfairness in the long term—even if fairness considerations were taken into account in the policy design process. In this paper, we propose a novel framework for studying long-term group fairness in dynamical systems, in which current decisions may affect an individual’s features in the next step, and thus, future decisions. Specifically, our framework allows us to identify a time-independent policy that converges, if deployed, to the targeted fair stationary state of the system in the long-term, independently of the initial data distribution. We model the system dynamics with a time-homogeneous Markov chain and optimize the policy leveraging the Markov Chain Convergence Theorem to ensure unique convergence. Our framework enables the utilization of historical temporal data to tackle challenges associated with delayed feedback when learning long-term fair policies in practice. Importantly, our framework shows that interventions on the data distribution (e.g., subsidies) can be used to achieve policy learning that is both short- and long-term fair. We provide examples of different targeted fair states of the system, encompassing a range of long-term goals for society and policymakers. In semi-synthetic simulations based on real-world datasets, we show how our approach facilitates identifying effective interventions for long-term fairness.