Academic literature on the topic 'Fairness-Accuracy trade-Off'

The fairness in machine learning is getting increasing attention, as its applications in different fields continue to expand and diversify. To mitigate the discriminated model behaviors between different demographic groups, we introduce a novel post-processing method to optimize over multiple fairness constraints through group-aware threshold adaptation. We propose to learn adaptive classification thresholds for each demographic group by optimizing the confusion matrix estimated from the probability distribution of a classification model output. As we only need an estimated probability distribution of model output instead of the classification model structure, our post-processing model can be applied to a wide range of classification models and improve fairness in a model-agnostic manner and ensure privacy. This even allows us to post-process existing fairness methods to further improve the trade-off between accuracy and fairness. Moreover, our model has low computational cost. We provide rigorous theoretical analysis on the convergence of our optimization algorithm and the trade-off between accuracy and fairness. Our method theoretically enables a better upper bound in near optimality than previous method under the same condition. Experimental results demonstrate that our method outperforms state-of-the-art methods and obtains the result that is closest to the theoretical accuracy-fairness trade-off boundary.

Most research on fairness in Machine Learning assumes the relationship between fairness and accuracy to be a trade-off, with an increase in fairness leading to an unavoidable loss of accuracy. In this study, several approaches for fair Machine Learning are studied to experimentally analyze the relationship between accuracy and group fairness. The results indicated that group fairness and accuracy may even benefit each other, which emphasizes the importance of selecting appropriate measures for performance evaluation. This work provides a foundation for further studies on the adequate objectives of Machine Learning in the context of fair automated decision making.

Biased data can lead to unfair machine learning models, highlighting the importance of embedding fairness at the beginning of data analysis, particularly during dataset curation and labeling. In response, we propose Falcon, a scalable fair active learning framework. Falcon adopts a data-centric approach that improves machine learning model fairness via strategic sample selection. Given a user-specified group fairness measure, Falcon identifies samples from "target groups" (e.g., (attribute=female, label=positive)) that are the most informative for improving fairness. However, a challenge arises since these target groups are defined using ground truth labels that are not available during sample selection. To handle this, we propose a novel trial-and-error method, where we postpone using a sample if the predicted label is different from the expected one and falls outside the target group. We also observe the trade-off that selecting more informative samples results in higher likelihood of postponing due to undesired label prediction, and the optimal balance varies per dataset. We capture the trade-off between informativeness and postpone rate as policies and propose to automatically select the best policy using adversarial multi-armed bandit methods, given their computational efficiency and theoretical guarantees. Experiments show that Falcon significantly outperforms existing fair active learning approaches in terms of fairness and accuracy and is more efficient. In particular, only Falcon supports a proper trade-off between accuracy and fairness where its maximum fairness score is 1.8--4.5x higher than the second-best results.

As Federated Learning (FL) gains prominence in distributed machine learning applications, achieving fairness without compromising predictive performance becomes paramount. The data being gathered from distributed clients in an FL environment often leads to class imbalance. In such scenarios, balanced accuracy rather than accuracy is the true representation of model performance. However, most state-of-the-art fair FL methods report accuracy as the measure of performance, which can lead to misguided interpretations of the model's effectiveness to mitigate discrimination. To the best of our knowledge, this work presents the first attempt towards achieving Pareto-optimal trade-offs between balanced accuracy and fairness in a federated environment (FairTrade). By utilizing multi-objective optimization, the framework negotiates the intricate balance between model's balanced accuracy and fairness. The framework's agnostic design adeptly accommodates both statistical and causal fairness notions, ensuring its adaptability across diverse FL contexts. We provide empirical evidence of our framework's efficacy through extensive experiments on five real-world datasets and comparisons with six baselines. The empirical results underscore the potential of our framework in improving the trade-off between fairness and balanced accuracy in FL applications.

Accuracy and individual fairness are both crucial for trustworthy machine learning, but these two aspects are often incompatible with each other so that enhancing one aspect may sacrifice the other inevitably with side effects of true bias or false fairness. We propose in this paper a new fairness criterion, accurate fairness, to align individual fairness with accuracy. Informally, it requires the treatments of an individual and the individual's similar counterparts to conform to a uniform target, i.e., the ground truth of the individual. We prove that accurate fairness also implies typical group fairness criteria over a union of similar sub-populations. We then present a Siamese fairness in-processing approach to minimize the accuracy and fairness losses of a machine learning model under the accurate fairness constraints. To the best of our knowledge, this is the first time that a Siamese approach is adapted for bias mitigation. We also propose fairness confusion matrix-based metrics, fair-precision, fair-recall, and fair-F1 score, to quantify a trade-off between accuracy and individual fairness. Comparative case studies with popular fairness datasets show that our Siamese fairness approach can achieve on average 1.02%-8.78% higher individual fairness (in terms of fairness through awareness) and 8.38%-13.69% higher accuracy, as well as 10.09%-20.57% higher true fair rate, and 5.43%-10.01% higher fair-F1 score, than the state-of-the-art bias mitigation techniques. This demonstrates that our Siamese fairness approach can indeed improve individual fairness without trading accuracy. Finally, the accurate fairness criterion and Siamese fairness approach are applied to mitigate the possible service discrimination with a real Ctrip dataset, by on average fairly serving 112.33% more customers (specifically, 81.29% more customers in an accurately fair way) than baseline models.

We propose a general approach to fairness based on transporting distributions corresponding to different sensitive attributes to a common distribution. We use optimal transport theory to derive target distributions and methods that allow us to achieve fairness with minimal changes to the unfair model. Our approach is applicable to both classification and regression problems, can enforce different notions of fairness, and enable us to achieve a Pareto-optimal trade-off between accuracy and fairness. We demonstrate that it outperforms previous approaches in several benchmark fairness datasets.

One of the main concerns about fairness in machine learning (ML) is that, in order to achieve it, one may have to trade off some accuracy. To overcome this issue, Hardt et al. proposed the notion of equality of opportunity (EO), which is compatible with maximal accuracy when the target label is deterministic with respect to the input features. In the probabilistic case, however, the issue is more complicated: It has been shown that under differential privacy constraints, there are data sources for which EO can only be achieved at the total detriment of accuracy, in the sense that a classifier that satisfies EO cannot be more accurate than a trivial (random guessing) classifier. In our paper we strengthen this result by removing the privacy constraint. Namely, we show that for certain data sources, the most accurate classifier that satisfies EO is a trivial classifier. Furthermore, we study the trade-off between accuracy and EO loss (opportunity difference), and provide a sufficient condition on the data source under which EO and non-trivial accuracy are compatible.

Machine learning (ML) models are increasingly being used for high-stake applications that can greatly impact people’s lives. Sometimes, these models can be biased toward certain social groups on the basis of race, gender, or ethnicity. Many prior works have attempted to mitigate this “model discrimination” by updating the training data (pre-processing), altering the model learning process (in-processing), or manipulating the model output (post-processing). However, more work can be done in extending this situation to intersectional fairness, where we consider multiple sensitive parameters (e.g., race) and sensitive options (e.g., black or white), thus allowing for greater real-world usability. Prior work in fairness has also suffered from an accuracy–fairness trade-off that prevents both accuracy and fairness from being high. Moreover, the previous literature has not clearly presented holistic fairness metrics that work with intersectional fairness. In this paper, we address all three of these problems by (a) creating a bias mitigation technique called DualFair and (b) developing a new fairness metric (i.e., AWI, a measure of bias of an algorithm based upon inconsistent counterfactual predictions) that can handle intersectional fairness. Lastly, we test our novel mitigation method using a comprehensive U.S. mortgage lending dataset and show that our classifier, or fair loan predictor, obtains relatively high fairness and accuracy metrics.

Organizations that collect and sell data face increasing scrutiny for the discriminatory use of data. We propose a novel unsupervised approach to map data into a compressed binary representation independent of sensitive attributes. We show that in an information bottleneck framework, a parsimonious representation should filter out information related to sensitive attributes if they are provided directly to the decoder. Empirical results show that the method achieves state-of-the-art accuracy-fairness trade-off and that explicit control of the entropy of the representation bit stream allows the user to move smoothly and simultaneously along both rate-distortion and rate-fairness curves.

At present, gradient boosting decision trees (GBDTs) has become a popular machine learning algorithm and has shined in many data mining competitions and real-world applications for its salient results on classification, ranking, prediction, etc. Federated learning which aims to mitigate privacy risks and costs, enables many entities to keep data locally and train a model collaboratively under an orchestration service. However, most of the existing systems often fail to make an excellent trade-off between accuracy and communication. In addition, they overlook an important aspect: fairness such as performance gains from different parties’ datasets. In this paper, we propose a novel federated GBDT scheme based on the blockchain which can achieve constant communication overhead and good model performance and quantify the contribution of each party. Specifically, we replace the tree-based communication scheme with the pure gradient-based scheme and compress the intermediate gradient information to a limit to achieve good model performance and constant communication overhead in skewed datasets. On the other hand, we introduce a novel contribution allocation scheme named split Shapley value, which can quantify the contribution of each party with a limited gradient update and provide a basis for monetary reward. Finally, we combine the quantification mechanism with blockchain organically and implement a closed-loop federated GBDT system FGBDT-Chain in a permissioned blockchain environment and conduct a comprehensive experiment on public datasets. The experimental results show that FGBDT-Chain achieves a good trade-off between accuracy, communication overhead, fairness, and security under large-scale skewed datasets.

Les décisions algorithmiques sont actuellement utilisées quotidiennement. Ces décisions reposent souvent sur des algorithmes d'apprentissage automatique (machine learning, ML) qui peuvent produire des modèles complexes et opaques. Des études récentes ont soulevé des problèmes d'iniquité en révélant des résultats discriminatoires produits par les modèles ML contre des minorités et des groupes non privilégiés. Comme les modèles ML sont capables d'amplifier la discrimination en raison de résultats injustes, cela révèle la nécessité d'approches qui découvrent et suppriment les biais inattendues. L'évaluation de l'équité et l'atténuation de l'iniquité sont les deux tâches principales qui ont motivé la croissance du domaine de recherche en équité algorithmique (algorithmic fairness). Plusieurs notions utilisées pour évaluer l'équité se concentrent sur les résultats et sont liées à des attributs sensibles (par exemple, l'éthinicité) par des mesures statistiques. Bien que ces notions aient une sémantique distincte, l'utilisation de ces définitions est critiquée pour sa compréhension réductrice de l'équité, dont le but est essentiellement de mettre en œuvre des rapports d'acceptation/non-acceptation, ignorant d'autres perspectives sur l'iniquité et l'impact sociétal. Process fairness (équité des procédures) est au contraire une notion d'équité subjective, centrée sur le processus qui conduit aux résultats. Pour atténuer ou supprimer l'iniquité, les approches appliquent généralement des interventions en matière d'équité selon des étapes spécifiques. Elles modifient généralement soit les données avant l'apprentissage, la fonction d'optimisation ou les sorties des algorithmes afin d'obtenir des résultats plus équitables. Récemment, les recherches ont été consacrées à l'exploration de combinaisons de différentes interventions en matière d'équité, ce qui est désigné dans cette thèse par le traitement hybride de l'équité. Une fois que nous essayons d'atténuer l'iniquité, une tension entre l'équité et la performance apparaît, connue comme le compromis équité/précision. Cette thèse se concentre sur le problème du compromis équité/précision, puisque nous sommes intéressés par la réduction des biais inattendues sans compromettre les performances de classification. Nous proposons donc des méthodes ensemblistes pour trouver un bon compromis entre l'équité et la performance de classification des modèles ML, en particulier les classificateurs binaires. De plus, ces méthodes produisent des classificateurs d'ensemble grâce à une combinaison d'interventions sur l'équité, ce qui caractérise les approches de traitement hybride de l'équité. Nous proposons FixOut (FaIrness through eXplanations and feature dropOut), un framework centré sur l'humain et agnostique vis-à-vis des modèles qui améliore l'équité sans compromettre les performances de classification. Il reçoit en entrée un classificateur pré-entraîné, un ensemble de données, un ensemble de attributs sensibles et une méthode d'explication, et il produit un nouveau classificateur qui dépend moins des attributs sensibles. Pour évaluer la dépendance d'un modèle pré-entraîné aux attributs sensibles, FixOut utilise des explications pour estimer la contribution des attributs aux résultats du modèle. S'il s'avère que les attributs sensibles contribuent globalement aux résultats, alors le modèle est considéré comme injuste. Dans ce cas, il construit un groupe de classificateurs plus justes qui sont ensuite agrégés pour obtenir un modèle d'ensemble. Nous montrons l'adaptabilité de FixOut sur différentes combinaisons de méthodes d'explication et d'approches d'échantillonnage. Nous évaluons également l'efficacité de FixOut par rapport au process fairness mais aussi en utilisant des notions d'équité standard bien connues disponibles dans la littérature. De plus, nous proposons plusieurs améliorations telles que l'automatisation du choix des paramètres et l'extension de FixOut à d'autres types de données
Algorithmic decisions are currently being used on a daily basis. These decisions often rely on Machine Learning (ML) algorithms that may produce complex and opaque ML models. Recent studies raised unfairness concerns by revealing discriminating outcomes produced by ML models against minorities and unprivileged groups. As ML models are capable of amplifying discrimination against minorities due to unfair outcomes, it reveals the need for approaches that uncover and remove unintended biases. Assessing fairness and mitigating unfairness are the two main tasks that have motivated the growth of the research field called {algorithmic fairness}. Several notions used to assess fairness focus on the outcomes and link to sensitive features (e.g. gender and ethnicity) through statistical measures. Although these notions have distinct semantics, the use of these definitions of fairness is criticized for being a reductionist understanding of fairness whose aim is basically to implement accept/not-accept reports, ignoring other perspectives on inequality and on societal impact. Process fairness instead is a subjective fairness notion which is centered on the process that leads to outcomes. To mitigate or remove unfairness, approaches generally apply fairness interventions in specific steps. They usually change either (1) the data before training or (2) the optimization function or (3) the algorithms' outputs in order to enforce fairer outcomes. Recently, research on algorithmic fairness have been dedicated to explore combinations of different fairness interventions, which is referred to in this thesis as {fairness hybrid-processing}. Once we try to mitigate unfairness, a tension between fairness and performance arises that is known as the fairness-accuracy trade-off. This thesis focuses on the fairness-accuracy trade-off problem since we are interested in reducing unintended biases without compromising classification performance. We thus propose ensemble-based methods to find a good compromise between fairness and classification performance of ML models, in particular models for binary classification. In addition, these methods produce ensemble classifiers thanks to a combination of fairness interventions, which characterizes the fairness hybrid-processing approaches. We introduce FixOut ({F}a{I}rness through e{X}planations and feature drop{Out}), the human-centered, model-agnostic framework that improves process fairness without compromising classification performance. It receives a pre-trained classifier (original model), a dataset, a set of sensitive features, and an explanation method as input, and it outputs a new classifier that is less reliant on the sensitive features. To assess the reliance of a given pre-trained model on sensitive features, FixOut uses explanations to estimate the contribution of features to models' outcomes. If sensitive features are shown to contribute globally to models' outcomes, then the model is deemed unfair. In this case, it builds a pool of fairer classifiers that are then aggregated to obtain an ensemble classifier. We show the adaptability of FixOut on different combinations of explanation methods and sampling approaches. We also evaluate the effectiveness of FixOut w.r.t. to process fairness but also using well-known standard fairness notions available in the literature. Furthermore, we propose several improvements such as automating the choice of FixOut's parameters and extending FixOut to other data types

AbstractDecision trees are increasingly used to make socially sensitive decisions, where they are expected to be both accurate and fair, but it remains a challenging task to optimize the learning algorithm for fairness in a predictable and explainable fashion. To overcome the challenge, we propose an iterative framework for choosing decision attributes, or features, at each level by formulating feature selection as a series of mixed integer optimization problems. Both fairness and accuracy requirements are encoded as numerical constraints and solved by an off-the-shelf constraint solver. As a result, the trade-off between fairness and accuracy is quantifiable. At a high level, our method can be viewed as a generalization of the entropy-based greedy search techniques such as and , and existing fair learning techniques such as and . Our experimental evaluation on six datasets, for which demographic parity is used as the fairness metric, shows that the method is significantly more effective in reducing bias than other methods while maintaining accuracy. Furthermore, compared to non-iterative constraint solving, our iterative approach is at least 10 times faster.

Let's face it: We're all Web journalists now. You might be working for a newspaper or magazine, a television or radio outlet, but your story is still likely to end up on the Web as well as in its original medium. You or your publication may even provide supplemental material that appears only on the Web—say, a behind-the-scenes notebook, an interactive graphic, or a blog. Or you might even be a journalist whose work appears almost exclusively on the Web—like me. I worked at daily newspapers for 19 years before joining MSNBC, a combined Web/television news organization. So I still tend to think of the Web as an online newspaper, with a lot of text, some pictures, and a few extra twists. But with the passage of time, online journalism is gradually coming into its own—just as TV started out as radio with pictures, but soon became a distinct news medium. To my mind, the principles of online journalism—having to do with fairness, accuracy, and completeness—are the same as the principles of off-line journalism. But the medium does shape the message, as well as the qualities that each medium considers most important. Wire-service reporters value getting the story out fast; newspapers value exclusive sources; magazines value in-depth coverage; radio and TV look for sounds and pictures that will help tell the story. All these factors are important for the Web as well, but one thing makes online journalism unique: Web writers are looking for ways to tell the story using software. Let's take a closer look at how one multimedia story unfolded, then get into how the tools and toys of the trade can be used in your own work. News coverage of space shuttle launches and landings usually follows a familiar routine: From MSNBC's West Coast newsroom in Redmond, Washington, I would update the landing-day story continuously, starting with the de-orbit burn, just as a wire service reporter might do. On February 1, 2003, however, the shuttle's landing was scheduled for a Saturday morning, one of the lightest times of the week for Web traffic.