James Foulds
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
James Foulds has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Shimei Pan
Information Systems department, University of Maryland Baltimore County
16 shared publications - Rashidul Islam
Visa Research, USA
7 shared publications - Kamrun Naher Keya
University of Maryland, Baltimore County, USA
6 shared publications - Kamalika Chaudhuri
Dept. of Computer Science and Engineering University of California San Diego La Jolla CA
2 shared publications - Kiran Laxmikant Prabhu
Department of Information Systems, University of Maryland, Baltimore County (UMBC), 1000 Hilltop Cir., Baltimore, MD 21250, USA
2 shared publications - Max Welling
University of Amsterdam
2 shared publications
Latest Publications
- Open Access
- 2 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 211-219
When Fairness Metrics Help Each Other: Look-Ahead Optimization for Multi-Metric Fairness Weighting
Fairness-aware machine learning depends crucially on the fairness criterion chosen, and it is well understood that these criteria are often in conflict with each other. In this work, we challenge this conventional wisdom by exploring how competing fairness criteria may in some cases interact in helpful ways. We investigate whether adding an auxiliary fairness objective can improve a primary fairness objective while maintaining or improving predictive performance. We refer to such cases as fairness-metric synergies, and identify them using a Pareto improvement criterion. We first use fixed grid search to examine when these improvements appear across standard fairness metrics and socially consequential binary classification datasets, demonstrating that this phenomenon can indeed occur. To make this method more practical, we then propose LA-FW (Look-Ahead Fairness Weighting), a modular dynamic weighting algorithm that searches for useful fairness-weight configurations through short look-ahead updates during training. We analyze the computational relationship between LA-FW and fixed grid search and show that fixed grid search can be understood as a limiting case when candidate weights are evaluated with sufficient computation. Our experiments compare fixed grid search and LA-FW in terms of fairness, accuracy, and runtime, demonstrating substantial speedups while producing competitive fairness results across three datasets. Overall, this work adds to an emerging understanding that fairness metric trade-offs are more nuanced than was previously believed.
- 2 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 211-219
Open AccessWhen Fairness Metrics Help Each Other: Look-Ahead Optimization for Multi-Metric Fairness Weighting
Fairness-aware machine learning depends crucially on the fairness criterion chosen, and it is well understood that these criteria are often in conflict with each other. In this work, we challenge this conventional wisdom by exploring how competing fairness criteria may in some cases interact in helpful ways. We investigate whether adding an auxiliary fairness objective can improve a primary fairness objective while maintaining or improving predictive performance. We refer to such cases as fairness-metric synergies, and identify them using a Pareto improvement criterion. We first use fixed grid search to examine when these improvements appear across standard fairness metrics and socially consequential binary classification datasets, demonstrating that this phenomenon can indeed occur. To make this method more practical, we then propose LA-FW (Look-Ahead Fairness Weighting), a modular dynamic weighting algorithm that searches for useful fairness-weight configurations through short look-ahead updates during training. We analyze the computational relationship between LA-FW and fixed grid search and show that fixed grid search can be understood as a limiting case when candidate weights are evaluated with sufficient computation. Our experiments compare fixed grid search and LA-FW in terms of fairness, accuracy, and runtime, demonstrating substantial speedups while producing competitive fairness results across three datasets. Overall, this work adds to an emerging understanding that fairness metric trade-offs are more nuanced than was previously believed.
- Open Access
- 9 July 2026
- Published by IntechOpen
A Mandate Beyond Accuracy: Forging the Algorithm’s Social Contract
To realize the full promise of artificial intelligence (AI), its benefits must be shared equitably across all of society. A narrow focus on technical accuracy alone often creates inadvertent barriers, preventing technology from serving diverse populations. This chapter examines the systemic risks of an ‘accuracy-first’ development model, arguing that achieving social good within complex sociotechnical environments requires intentional trade-offs between raw performance and equitable outcomes. Central to this argument is the recognition that bias is not a singular event but a cumulative phenomenon embedded throughout the AI lifecycle. Because biases infiltrate every phase, from initial data semantics to final deploymen relying solely on post-hoc evaluation or isolated fairness metrics is insufficient. We argue for continuous lifecycle oversight to mitigate harms that otherwise cascade from individual and interpersonal levels to broader institutional, societal, and global environmental scales. As AI systems become omnipresent, this mindful approach to constant deployment is essential to prevent systemic exclusion. Finally, this chapter proposes a hopeful blueprint for a new social contract for AI. By embedding principles of fairness, building trust through explainable AI (XAI), and implementing robust, multi-level governance, we can transform AI from a mere tool of efficiency into a powerful catalyst for universal opportunity. Through this framework, we can build a future where technology empowers every community, ensuring that the evolution of AI leaves no one behind.
- 9 July 2026
- Published by IntechOpen
Open AccessA Mandate Beyond Accuracy: Forging the Algorithm’s Social Contract
To realize the full promise of artificial intelligence (AI), its benefits must be shared equitably across all of society. A narrow focus on technical accuracy alone often creates inadvertent barriers, preventing technology from serving diverse populations. This chapter examines the systemic risks of an ‘accuracy-first’ development model, arguing that achieving social good within complex sociotechnical environments requires intentional trade-offs between raw performance and equitable outcomes. Central to this argument is the recognition that bias is not a singular event but a cumulative phenomenon embedded throughout the AI lifecycle. Because biases infiltrate every phase, from initial data semantics to final deploymen relying solely on post-hoc evaluation or isolated fairness metrics is insufficient. We argue for continuous lifecycle oversight to mitigate harms that otherwise cascade from individual and interpersonal levels to broader institutional, societal, and global environmental scales. As AI systems become omnipresent, this mindful approach to constant deployment is essential to prevent systemic exclusion. Finally, this chapter proposes a hopeful blueprint for a new social contract for AI. By embedding principles of fairness, building trust through explainable AI (XAI), and implementing robust, multi-level governance, we can transform AI from a mere tool of efficiency into a powerful catalyst for universal opportunity. Through this framework, we can build a future where technology empowers every community, ensuring that the evolution of AI leaves no one behind.
- Open Access
- 16 May 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (5) , 398
A Comparative Simulation Study of the Fairness and Accuracy of Predictive Policing Systems in Baltimore City
There are ongoing discussions about predictive policing systems being unfair, for example, by exhibiting racial bias. Law enforcement in some cities, such as Los Angeles, California, and Baltimore, Maryland, have initiated the integration of these systems into their decision-making processes, and some of these systems were advertised as being unbiased. However, later studies discovered that these methods could also be unfair due to feedback loops and being trained on historically biased recorded data. Comparative studies on predictive policing systems are few and insufficiently comprehensive. Crucially, the relative fairness of predictive policing methods with regard to traditional hot spot-based policing has not been established. Moreover, the relationship between fairness and accuracy is complex and requires further study. Furthermore, the case of Baltimore City, Maryland, USA, has not yet been systematically analyzed despite its relevance as an early adopter of predictive policing technologies with a fraught history of social justice concerns around policing. An improved understanding of these questions could better inform policy decisions around predictive policing technologies both in Baltimore and beyond. Therefore, in this work we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than previously assumed. While we find that predictive policing exhibits bias due to feedback loops, as previously reported, we also find traditional hot spot-based policing to have similar issues. Although predictive policing is found to be more fair and accurate than hot spot policing in the short term, it also amplifies bias more quickly, suggesting the potential for worse long-run behavior. In Baltimore, the bias in these systems tended toward over-policing White neighborhoods in some cases, unlike in previous studies. However, when the analysis was restricted to some specific crime types, this tendency differed. Overall, this work demonstrates a methodology for city-specific evaluation and compares behavioral tendencies of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies. We recommend that authorities and community stakeholders use simulation methodologies to assist in collaboratively navigating the complexities around fairness in predictive policing.
- 16 May 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (5) , 398
Open AccessA Comparative Simulation Study of the Fairness and Accuracy of Predictive Policing Systems in Baltimore City
There are ongoing discussions about predictive policing systems being unfair, for example, by exhibiting racial bias. Law enforcement in some cities, such as Los Angeles, California, and Baltimore, Maryland, have initiated the integration of these systems into their decision-making processes, and some of these systems were advertised as being unbiased. However, later studies discovered that these methods could also be unfair due to feedback loops and being trained on historically biased recorded data. Comparative studies on predictive policing systems are few and insufficiently comprehensive. Crucially, the relative fairness of predictive policing methods with regard to traditional hot spot-based policing has not been established. Moreover, the relationship between fairness and accuracy is complex and requires further study. Furthermore, the case of Baltimore City, Maryland, USA, has not yet been systematically analyzed despite its relevance as an early adopter of predictive policing technologies with a fraught history of social justice concerns around policing. An improved understanding of these questions could better inform policy decisions around predictive policing technologies both in Baltimore and beyond. Therefore, in this work we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than previously assumed. While we find that predictive policing exhibits bias due to feedback loops, as previously reported, we also find traditional hot spot-based policing to have similar issues. Although predictive policing is found to be more fair and accurate than hot spot policing in the short term, it also amplifies bias more quickly, suggesting the potential for worse long-run behavior. In Baltimore, the bias in these systems tended toward over-policing White neighborhoods in some cases, unlike in previous studies. However, when the analysis was restricted to some specific crime types, this tendency differed. Overall, this work demonstrates a methodology for city-specific evaluation and compares behavioral tendencies of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies. We recommend that authorities and community stakeholders use simulation methodologies to assist in collaboratively navigating the complexities around fairness in predictive policing.
- Open Access
A Comparative Simulation Study of the Fairness and Accuracy of Predictive Policing Systems in Baltimore City
There are ongoing discussions about predictive policing systems, such as those deployed in Los Angeles, California and Baltimore, Maryland, being unfair, for example, by exhibiting racial bias. Studies found that unfairness may be due to feedback loops and being trained on historically biased recorded data. However, comparative studies on predictive policing systems are few and are not sufficiently comprehensive. In this work, we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than was previously assumed. While predictive policing exhibited bias due to feedback loops as was previously reported, we found that the traditional alternative, hot spots policing, had similar issues. Predictive policing was found to be more fair and accurate than hot spots policing in the short term, although it amplified bias faster, suggesting the potential for worse long-run behavior. In Baltimore, in some cases the bias in these systems tended toward over-policing in White neighborhoods, unlike in previous studies. Overall, this work demonstrates a methodology for city-specific evaluation and behavioral-tendency comparison of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies.
Open AccessA Comparative Simulation Study of the Fairness and Accuracy of Predictive Policing Systems in Baltimore City
There are ongoing discussions about predictive policing systems, such as those deployed in Los Angeles, California and Baltimore, Maryland, being unfair, for example, by exhibiting racial bias. Studies found that unfairness may be due to feedback loops and being trained on historically biased recorded data. However, comparative studies on predictive policing systems are few and are not sufficiently comprehensive. In this work, we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than was previously assumed. While predictive policing exhibited bias due to feedback loops as was previously reported, we found that the traditional alternative, hot spots policing, had similar issues. Predictive policing was found to be more fair and accurate than hot spots policing in the short term, although it amplified bias faster, suggesting the potential for worse long-run behavior. In Baltimore, in some cases the bias in these systems tended toward over-policing in White neighborhoods, unlike in previous studies. Overall, this work demonstrates a methodology for city-specific evaluation and behavioral-tendency comparison of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies.
- Open Access
A Unifying Human-Centered AI Fairness Framework
The increasing use of Artificial Intelligence (AI) in critical societal domains has amplified concerns about fairness, particularly regarding unequal treatment across sensitive attributes such as race, gender, and socioeconomic status. While there has been substantial work on ensuring AI fairness, navigating trade-offs between competing notions of fairness as well as predictive accuracy remains challenging, creating barriers to the practical deployment of fair AI systems. To address this, we introduce a unifying human-centered fairness framework that systematically covers eight distinct fairness metrics, formed by combining individual and group fairness, infra-marginal and intersectional assumptions, and outcome-based and equality-of-opportunity (EOO) perspectives. This structure allows stakeholders to align fairness interventions with their values and contextual considerations. The framework uses a consistent and easy-to-understand formulation for all metrics to reduce the learning curve for non-experts. Rather than privileging a single fairness notion, the framework enables stakeholders to assign weights across multiple fairness objectives, reflecting their priorities and facilitating multi-stakeholder compromises. We apply this approach to four real-world datasets: the UCI Adult census dataset for income prediction, the COMPAS dataset for criminal recidivism, the German Credit dataset for credit risk assessment, and the MEPS dataset for healthcare utilization. We show that adjusting weights reveals nuanced trade-offs between different fairness metrics. Finally, through case studies in judicial decision-making and healthcare, we demonstrate how the framework can inform practical and value-sensitive deployment of fair AI systems.
Open AccessA Unifying Human-Centered AI Fairness Framework
The increasing use of Artificial Intelligence (AI) in critical societal domains has amplified concerns about fairness, particularly regarding unequal treatment across sensitive attributes such as race, gender, and socioeconomic status. While there has been substantial work on ensuring AI fairness, navigating trade-offs between competing notions of fairness as well as predictive accuracy remains challenging, creating barriers to the practical deployment of fair AI systems. To address this, we introduce a unifying human-centered fairness framework that systematically covers eight distinct fairness metrics, formed by combining individual and group fairness, infra-marginal and intersectional assumptions, and outcome-based and equality-of-opportunity (EOO) perspectives. This structure allows stakeholders to align fairness interventions with their values and contextual considerations. The framework uses a consistent and easy-to-understand formulation for all metrics to reduce the learning curve for non-experts. Rather than privileging a single fairness notion, the framework enables stakeholders to assign weights across multiple fairness objectives, reflecting their priorities and facilitating multi-stakeholder compromises. We apply this approach to four real-world datasets: the UCI Adult census dataset for income prediction, the COMPAS dataset for criminal recidivism, the German Credit dataset for credit risk assessment, and the MEPS dataset for healthcare utilization. We show that adjusting weights reveals nuanced trade-offs between different fairness metrics. Finally, through case studies in judicial decision-making and healthcare, we demonstrate how the framework can inform practical and value-sensitive deployment of fair AI systems.
- Open Access
- 7 August 2025
- Published by IOS Press
- Vol. 329, 1049-1053
Facilitating Online Healthcare Support Group Formation Using Topic Modeling
Patients increasingly seek peer support in online health forums; however, the large-scale and inconsistent engagement patterns of these forums often fail to meet patients’ support needs effectively. Smaller, personalized support groups could address these challenges by tailoring interactions to users’ shared experiences and demographics. This study introduces the Group-specific Dirichlet Multinomial Regression (gDMR) model, a structured framework for automating and personalizing support group formation using user-generated content, demographic, and interaction data. By incorporating group-specific parameters and node embeddings, gDMR captures nuanced demographic and behavioral patterns, extending traditional Dirichlet Multinomial Regression (DMR). Experiments demonstrate gDMR’s ability to form groups that are more semantically coherent and contextually relevant than those produced by baseline models. This scalable model reduces manual effort in group personalization, fostering inclusive engagement and enhancing patient-centered care. Findings highlight gDMR’s potential as a framework for digital healthcare platforms, from online health communities to healthcare systems utilizing electronic health records (EHRs), advancing health informatics through support group formation.
- 7 August 2025
- Published by IOS Press
- Vol. 329, 1049-1053
Open AccessFacilitating Online Healthcare Support Group Formation Using Topic Modeling
Patients increasingly seek peer support in online health forums; however, the large-scale and inconsistent engagement patterns of these forums often fail to meet patients’ support needs effectively. Smaller, personalized support groups could address these challenges by tailoring interactions to users’ shared experiences and demographics. This study introduces the Group-specific Dirichlet Multinomial Regression (gDMR) model, a structured framework for automating and personalizing support group formation using user-generated content, demographic, and interaction data. By incorporating group-specific parameters and node embeddings, gDMR captures nuanced demographic and behavioral patterns, extending traditional Dirichlet Multinomial Regression (DMR). Experiments demonstrate gDMR’s ability to form groups that are more semantically coherent and contextually relevant than those produced by baseline models. This scalable model reduces manual effort in group personalization, fostering inclusive engagement and enhancing patient-centered care. Findings highlight gDMR’s potential as a framework for digital healthcare platforms, from online health communities to healthcare systems utilizing electronic health records (EHRs), advancing health informatics through support group formation.
- Open Access
You've Changed: Detecting Modification of Black-Box Large Language Models
Large Language Models (LLMs) are often provided as a service via an API, making it challenging for developers to detect changes in their behavior. We present an approach to monitor LLMs for changes by comparing the distributions of linguistic and psycholinguistic features of generated text. Our method uses a statistical test to determine whether the distributions of features from two samples of text are equivalent, allowing developers to identify when an LLM has changed. We demonstrate the effectiveness of our approach using five OpenAI completion models and Meta's Llama 3 70B chat model. Our results show that simple text features coupled with a statistical test can distinguish between language models. We also explore the use of our approach to detect prompt injection attacks. Our work enables frequent LLM change monitoring and avoids computationally expensive benchmark evaluations.
Open AccessYou've Changed: Detecting Modification of Black-Box Large Language Models
Large Language Models (LLMs) are often provided as a service via an API, making it challenging for developers to detect changes in their behavior. We present an approach to monitor LLMs for changes by comparing the distributions of linguistic and psycholinguistic features of generated text. Our method uses a statistical test to determine whether the distributions of features from two samples of text are equivalent, allowing developers to identify when an LLM has changed. We demonstrate the effectiveness of our approach using five OpenAI completion models and Meta's Llama 3 70B chat model. Our results show that simple text features coupled with a statistical test can distinguish between language models. We also explore the use of our approach to detect prompt injection attacks. Our work enables frequent LLM change monitoring and avoids computationally expensive benchmark evaluations.
- Open Access
LLM-based Corroborating and Refuting Evidence Retrieval for Scientific Claim Verification
In this paper, we introduce CIBER (Claim Investigation Based on Evidence Retrieval), an extension of the Retrieval-Augmented Generation (RAG) framework designed to identify corroborating and refuting documents as evidence for scientific claim verification. CIBER addresses the inherent uncertainty in Large Language Models (LLMs) by evaluating response consistency across diverse interrogation probes. By focusing on the behavioral analysis of LLMs without requiring access to their internal information, CIBER is applicable to both white-box and black-box models. Furthermore, CIBER operates in an unsupervised manner, enabling easy generalization across various scientific domains. Comprehensive evaluations conducted using LLMs with varying levels of linguistic proficiency reveal CIBER's superior performance compared to conventional RAG approaches. These findings not only highlight the effectiveness of CIBER but also provide valuable insights for future advancements in LLM-based scientific claim verification.
Open AccessLLM-based Corroborating and Refuting Evidence Retrieval for Scientific Claim Verification
In this paper, we introduce CIBER (Claim Investigation Based on Evidence Retrieval), an extension of the Retrieval-Augmented Generation (RAG) framework designed to identify corroborating and refuting documents as evidence for scientific claim verification. CIBER addresses the inherent uncertainty in Large Language Models (LLMs) by evaluating response consistency across diverse interrogation probes. By focusing on the behavioral analysis of LLMs without requiring access to their internal information, CIBER is applicable to both white-box and black-box models. Furthermore, CIBER operates in an unsupervised manner, enabling easy generalization across various scientific domains. Comprehensive evaluations conducted using LLMs with varying levels of linguistic proficiency reveal CIBER's superior performance compared to conventional RAG approaches. These findings not only highlight the effectiveness of CIBER but also provide valuable insights for future advancements in LLM-based scientific claim verification.
- Open Access
- 4 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 526-535
DoubleDistillation: Enhancing LLMs for Informal Text Analysis using Multistage Knowledge Distillation from Speech and Text
Traditional large language models (LLMs) leverage extensive text corpora but lack access to acoustic and para-linguistic cues present in speech. There is a growing interest in enhancing text-based models with audio information. However, current models often require an aligned audio-text dataset which is frequently much smaller than typical language model training corpora. Moreover, these models often require both text and audio streams during inference/testing. In this study, we introduce a novel two-stage knowledge distillation (KD) approach that enables language models to (a) incorporate rich acoustic and paralinguistic information from speech, (b) utilize text corpora comparable in size to typical language model training data, and (c) support text-only analysis without requiring an audio stream during inference/testing. Specifically, we employ a pre-trained speech embedding teacher model (OpenAI Whisper) to train a Teacher Assistant (TA) model on an aligned audio-text dataset in the first stage. In the second stage, the TA’s knowledge is transferred to a student language model trained on a conventional text dataset. Thus, our two-stage KD method leverages both the acoustic and paralinguistic cues in the aligned audio-text data and the nuanced linguistic knowledge in a large text-only dataset. Based on our evaluation, this DoubleDistillation system consistently outperforms traditional LLMs in 15 informal text understanding tasks.
- 4 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 526-535
Open AccessDoubleDistillation: Enhancing LLMs for Informal Text Analysis using Multistage Knowledge Distillation from Speech and Text
Traditional large language models (LLMs) leverage extensive text corpora but lack access to acoustic and para-linguistic cues present in speech. There is a growing interest in enhancing text-based models with audio information. However, current models often require an aligned audio-text dataset which is frequently much smaller than typical language model training corpora. Moreover, these models often require both text and audio streams during inference/testing. In this study, we introduce a novel two-stage knowledge distillation (KD) approach that enables language models to (a) incorporate rich acoustic and paralinguistic information from speech, (b) utilize text corpora comparable in size to typical language model training data, and (c) support text-only analysis without requiring an audio stream during inference/testing. Specifically, we employ a pre-trained speech embedding teacher model (OpenAI Whisper) to train a Teacher Assistant (TA) model on an aligned audio-text dataset in the first stage. In the second stage, the TA’s knowledge is transferred to a student language model trained on a conventional text dataset. Thus, our two-stage KD method leverages both the acoustic and paralinguistic cues in the aligned audio-text data and the nuanced linguistic knowledge in a large text-only dataset. Based on our evaluation, this DoubleDistillation system consistently outperforms traditional LLMs in 15 informal text understanding tasks.
- Open Access
- 4 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 88-92
Towards A Unifying Human-Centered AI Fairness Framework
Achieving fairness in AI systems is a critical yet challenging task due to conflicting metrics and their underlying societal assumptions, e.g., the extent to which racist and sexist societal processes are presumed to cause harm and the extent to which we should apply affirmative corrections. Moreover, these measures often contradict each other and might also make the AI system less accurate. This work takes a step towards a unifying human-centered fairness framework to guide stakeholders in navigating these complexities, including their potential incompatibility and the corresponding trade-offs. Our framework acknowledges the spectrum of fairness definitions —individual vs. group fairness, infra-marginal (politically conservative) vs. intersectional (politically progressive) treatment of disparities— allowing stakeholders to prioritize desired outcomes by assigning weights to various fairness considerations, trading them off against each other, as well as predictive performance, supporting stakeholders in exploring the impacts of their fairness choices to achieve a consensus solution. Our learning algorithms then ensure the resulting AI system reflects the stakeholder-chosen priorities. By enabling multi-stakeholder compromises, our framework can potentially mitigate individual analysts’ subjectivity. We performed experiments to validate our methods on the UCI Adult census dataset and the COMPAS criminal recidivism dataset.
- 4 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 88-92
Open AccessTowards A Unifying Human-Centered AI Fairness Framework
Achieving fairness in AI systems is a critical yet challenging task due to conflicting metrics and their underlying societal assumptions, e.g., the extent to which racist and sexist societal processes are presumed to cause harm and the extent to which we should apply affirmative corrections. Moreover, these measures often contradict each other and might also make the AI system less accurate. This work takes a step towards a unifying human-centered fairness framework to guide stakeholders in navigating these complexities, including their potential incompatibility and the corresponding trade-offs. Our framework acknowledges the spectrum of fairness definitions —individual vs. group fairness, infra-marginal (politically conservative) vs. intersectional (politically progressive) treatment of disparities— allowing stakeholders to prioritize desired outcomes by assigning weights to various fairness considerations, trading them off against each other, as well as predictive performance, supporting stakeholders in exploring the impacts of their fairness choices to achieve a consensus solution. Our learning algorithms then ensure the resulting AI system reflects the stakeholder-chosen priorities. By enabling multi-stakeholder compromises, our framework can potentially mitigate individual analysts’ subjectivity. We performed experiments to validate our methods on the UCI Adult census dataset and the COMPAS criminal recidivism dataset.