skip to main content

Title: The Fairness of Risk Scores Beyond Classification: Bipartite Ranking and the XAUC Metric
Where machine-learned predictive risk scores inform high-stakes decisions, such as bail and sentencing in criminal justice, fairness has been a serious concern. Recent work has characterized the disparate impact that such risk scores can have when used for a binary classification task. This may not account, however, for the more diverse downstream uses of risk scores and their non-binary nature. To better account for this, in this paper, we investigate the fairness of predictive risk scores from the point of view of a bipartite ranking task, where one seeks to rank positive examples higher than negative ones. We introduce the xAUC disparity as a metric to assess the disparate impact of risk scores and define it as the difference in the probabilities of ranking a random positive example from one protected group above a negative one from another group and vice versa. We provide a decomposition of bipartite ranking loss into components that involve the discrepancy and components that involve pure predictive ability within each group. We use xAUC analysis to audit predictive risk scores for recidivism prediction, income prediction, and cardiac arrest prediction, where it describes disparities that are not evident from simply comparing within-group predictive performance.  more » « less
Award ID(s):
Author(s) / Creator(s):
Date Published:
Journal Name:
Advances in neural information processing systems
Page Range / eLocation ID:
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Many selection procedures involve ordering candidates according to their qualifications. For example, a university might order applicants according to a perceived probability of graduation within four years, and then select the top 1000 applicants. In this work, we address the problem of ranking members of a population according to their “probability” of success, based on a training set of historical binary outcome data (e.g., graduated in four years or not). We show how to obtain rankings that satisfy a number of desirable accuracy and fairness criteria, despite the coarseness of the training data. As the task of ranking is global (the rank of every individual depends not only on their own qualifications, but also on every other individuals’ qualifications), ranking is more subtle and vulnerable to manipulation than standard prediction tasks. Towards mitigating unfair discrimination caused by inaccuracies in rankings, we develop two parallel definitions of evidence-based rankings. The first definition relies on a semantic notion of domination-compatibility: if the training data suggest that members of a set S are more qualified (on average) than the members of T, then a ranking that favors T over S (where T dominates S) is blatantly inconsistent with the evidence, and likely to be discriminatory. The definition asks for domination-compatibility, not just for a pair of sets, but rather for every pair of sets from a rich collection C of subpopulations. The second definition aims at precluding even more general forms of discrimination; this notion of evidence-consistency requires that the ranking must be justified on the basis of consistency with the expectations for every set in the collection C. Somewhat surprisingly, while evidence-consistency is a strictly stronger notion than domination-compatibility when the collection C is predefined, the two notions are equivalent when the collection C may depend on the ranking in question. 
    more » « less
  2. A variety of fairness constraints have been proposed in the literature to mitigate group-level statistical bias. Their impacts have been largely evaluated for different groups of populations corresponding to a set of sensitive attributes, such as race or gender. Nonetheless, the community has not observed sufficient explorations for how imposing fairness constraints fare at an instance level. Building on the concept of influence function, a measure that characterizes the impact of a training example on the target model and its predictive performance, this work studies the influence of training examples when fairness constraints are imposed. We find out that under certain assumptions, the influence function with respect to fairness constraints can be decomposed into a kernelized combination of training examples. One promising application of the proposed fairness influence function is to identify suspicious training examples that may cause model discrimination by ranking their influence scores. We demonstrate with extensive experiments that training on a subset of weighty data examples leads to lower fairness violations with a trade-off of accuracy. 
    more » « less
  3. null (Ed.)
    Abstract Accurate prediction of suicide risk among children and adolescents within an actionable time frame is an important but challenging task. Very few studies have comprehensively considered the clinical risk factors available to produce quantifiable risk scores for estimation of short- and long-term suicide risk for pediatric population. In this paper, we built machine learning models for predicting suicidal behavior among children and adolescents based on their longitudinal clinical records, and determining short- and long-term risk factors. This retrospective study used deidentified structured electronic health records (EHR) from the Connecticut Children’s Medical Center covering the period from 1 October 2011 to 30 September 2016. Clinical records of 41,721 young patients (10–18 years old) were included for analysis. Candidate predictors included demographics, diagnosis, laboratory tests, and medications. Different prediction windows ranging from 0 to 365 days were adopted. For each prediction window, candidate predictors were first screened by univariate statistical tests, and then a predictive model was built via a sequential forward feature selection procedure. We grouped the selected predictors and estimated their contributions to risk prediction at different prediction window lengths. The developed predictive models predicted suicidal behavior across all prediction windows with AUCs varying from 0.81 to 0.86. For all prediction windows, the models detected 53–62% of suicide-positive subjects with 90% specificity. The models performed better with shorter prediction windows and predictor importance varied across prediction windows, illustrating short- and long-term risks. Our findings demonstrated that routinely collected EHRs can be used to create accurate predictive models for suicide risk among children and adolescents. 
    more » « less
  4. null (Ed.)
    Testing multiple subjects within a group, with a single test applied to the group (i.e., group testing), is an important tool for classifying populations as positive or negative for a specific binary characteristic in an efficient manner. We study the design of easily implementable, static group testing schemes that take into account operational constraints, heterogeneous populations, and uncertainty in subject risk, while considering classification accuracy- and robustness-based objectives. We derive key structural properties of optimal risk-based designs and show that the problem can be formulated as network flow problems. Our reformulation involves computationally expensive high-dimensional integrals. We develop an analytical expression that eliminates the need to compute high-dimensional integrals, drastically improving the tractability of constructing the underlying network. We demonstrate the impact through a case study on chlamydia screening, which leads to the following insights: (1) Risk-based designs are shown to be less expensive, more accurate, and more robust than current practices. (2) The performance of static risk-based schemes comprised of only two group sizes is comparable to those comprised of many group sizes. (3) Static risk-based schemes are an effective alternative to more complicated dynamic schemes. (4) An expectation-based formulation captures almost all benefits of a static risk-based scheme. 
    more » « less
  5. In social choice, traditional Kemeny rank aggregation combines the preferences of voters, expressed as rankings, into a single consensus ranking without consideration for how this ranking may unfairly affect marginalized groups (i.e., racial or gender). Developing fair rank aggregation methods is critical due to their societal influence in applications prioritizing job applicants, funding proposals, and scheduling medical patients. In this work, we introduce the Fair Exposure Kemeny Aggregation Problem (FairExp-kap) for combining vast and diverse voter preferences into a single ranking that is not only a suitable consensus, but ensures opportunities are not withheld from marginalized groups. In formalizing FairExp-kap, we extend the fairness of exposure notion from information retrieval to the rank aggregation context and present a complimentary metric for voter preference representation. We design algorithms for solving FairExp-kap that explicitly account for position bias, a common ranking-based concern that end-users pay more attention to higher ranked candidates. epik solves FairExp-kap exactly by incorporating non-pairwise fairness of exposure into the pairwise Kemeny optimization; while the approximate epira is a candidate swapping algorithm, that guarantees ranked candidate fairness. Utilizing comprehensive synthetic simulations and six real-world datasets, we show the efficacy of our approach illustrating that we succeed in mitigating disparate group exposure unfairness in consensus rankings, while maximally representing voter preferences. 
    more » « less