skip to main content

Title: Optimized Score Transformation for Fair Classification
This paper considers fair probabilistic classification where the outputs of primary interest are predicted probabilities, commonly referred to as scores. We formulate the problem of transforming scores to satisfy fairness constraints while minimizing the loss in utility. The formulation can be applied either to post-process classifier outputs or to pre-process training data, thus allowing maximum freedom in selecting a classification algorithm. We derive a closed-form expression for the optimal transformed scores and a convex optimization problem for the transformation parameters. In the population limit, the transformed score function is the fairness-constrained minimizer of cross-entropy with respect to the optimal unconstrained scores. In the finite sample setting, we propose to approach this solution using a combination of standard probabilistic classifiers and ADMM. Comprehensive experiments comparing to 10 existing methods show that the proposed FairScoreTransformer has advantages for score-based metrics such as Brier score and AUC while remaining competitive for binary label-based metrics such as accuracy.  more » « less
Award ID(s):
Author(s) / Creator(s):
; ;
Date Published:
Journal Name:
Proceedings of Machine Learning Research
Page Range / eLocation ID:
Medium: X
Sponsoring Org:
National Science Foundation
More Like this

    Operational earthquake forecasting for risk management and communication during seismic sequences depends on our ability to select an optimal forecasting model. To do this, we need to compare the performance of competing models in prospective experiments, and to rank their performance according to the outcome using a fair, reproducible and reliable method, usually in a low-probability environment. The Collaboratory for the Study of Earthquake Predictability conducts prospective earthquake forecasting experiments around the globe. In this framework, it is crucial that the metrics used to rank the competing forecasts are ‘proper’, meaning that, on average, they prefer the data generating model. We prove that the Parimutuel Gambling score, proposed, and in some cases applied, as a metric for comparing probabilistic seismicity forecasts, is in general ‘improper’. In the special case where it is proper, we show it can still be used improperly. We demonstrate the conclusions both analytically and graphically providing a set of simulation based techniques that can be used to assess if a score is proper or not. They only require a data generating model and, at least two forecasts to be compared. We compare the Parimutuel Gambling score’s performance with two commonly used proper scores (the Brier and logarithmic scores) using confidence intervals to account for the uncertainty around the observed score difference. We suggest that using confidence intervals enables a rigorous approach to distinguish between the predictive skills of candidate forecasts, in addition to their rankings. Our analysis shows that the Parimutuel Gambling score is biased, and the direction of the bias depends on the forecasts taking part in the experiment. Our findings suggest the Parimutuel Gambling score should not be used to distinguishing between multiple competing forecasts, and for care to be taken in the case where only two are being compared.

    more » « less
  2. Abstract Statistical relational learning (SRL) frameworks are effective at defining probabilistic models over complex relational data. They often use weighted first-order logical rules where the weights of the rules govern probabilistic interactions and are usually learned from data. Existing weight learning approaches typically attempt to learn a set of weights that maximizes some function of data likelihood; however, this does not always translate to optimal performance on a desired domain metric, such as accuracy or F1 score. In this paper, we introduce a taxonomy of search-based weight learning approaches for SRL frameworks that directly optimize weights on a chosen domain performance metric. To effectively apply these search-based approaches, we introduce a novel projection, referred to as scaled space (SS), that is an accurate representation of the true weight space. We show that SS removes redundancies in the weight space and captures the semantic distance between the possible weight configurations. In order to improve the efficiency of search, we also introduce an approximation of SS which simplifies the process of sampling weight configurations. We demonstrate these approaches on two state-of-the-art SRL frameworks: Markov logic networks and probabilistic soft logic. We perform empirical evaluation on five real-world datasets and evaluate them each on two different metrics. We also compare them against four other weight learning approaches. Our experimental results show that our proposed search-based approaches outperform likelihood-based approaches and yield up to a 10% improvement across a variety of performance metrics. Further, we perform an extensive evaluation to measure the robustness of our approach to different initializations and hyperparameters. The results indicate that our approach is both accurate and robust. 
    more » « less
  3. Deaf and hard of hearing individuals regularly rely on captioning while watching live TV. Live TV captioning is evaluated by regulatory agencies using various caption evaluation metrics. However, caption evaluation metrics are often not informed by preferences of DHH users or how meaningful the captions are. There is a need to construct caption evaluation metrics that take the relative importance of words in transcript into account. We conducted correlation analysis between two types of word embeddings and human-annotated labelled word-importance scores in existing corpus. We found that normalized contextualized word embeddings generated using BERT correlated better with manually annotated importance scores than word2vec-based word embeddings. We make available a pairing of word embeddings and their human-annotated importance scores. We also provide proof-of-concept utility by training word importance models, achieving an F1-score of 0.57 in the 6-class word importance classification task. 
    more » « less
  4. Abstract Background

    The 2021 NIJ recidivism forecasting challenge asks participants to construct predictive models of recidivism while balancing false positive rates across groups of Black and white individuals through a multiplicative fairness score. We investigate the performance of several models for forecasting 1-year recidivism and optimizing the NIJ multiplicative fairness metric.


    We consider standard linear and logistic regression, a penalized regression that optimizes a convex surrogate loss (that we show has an analytical solution), two post-processing techniques, linear regression with re-balanced data, a black-box general purpose optimizer applied directly to the NIJ metric and a gradient boosting machine learning approach.


    For the set of models investigated, we find that a simple heuristic of truncating scores at the decision threshold (thus predicting no recidivism across the data) yields as good or better NIJ fairness scores on held out data compared to other, more sophisticated approaches. We also find that when the cutoff is further away from the base rate of recidivism, as is the case in the competition where the base rate is 0.29 and the cutoff is 0.5, then simply optimizing the mean square error gives nearly optimal NIJ fairness metric solutions.


    The multiplicative metric in the 2021 NIJ recidivism forecasting competition encourages solutions that simply optimize MSE and/or use truncation, therefore yielding trivial solutions that forecast no one will recidivate.

    more » « less
  5. Subset selection is an integral component of AI systems that is increasingly affecting people’s livelihoods in applications ranging from hiring, healthcare, education, to financial decisions. Subset selections powered by AI-based methods include top- analytics, data summarization, clustering, and multi-winner voting. While group fairness auditing tools have been proposed for classification systems, these state-of-the-art tools are not directly applicable to measuring and conceptualizing fairness in selected subsets. In this work, we introduce the first comprehensive auditing framework, FINS, to support stakeholders in interpretably quantifying group fairness across a diverse range of subset-specific fairness concerns. FINS offers a family of novel measures that provide a flexible means to audit group fairness for fairness goals ranging from item-based, score-based, and a combination thereof. FINS provides one unified easy-to-understand interpretation across these different fairness problems. Further, we develop guidelines through the FINS Fair Subset Chart, that supports auditors in determining which measures are relevant to their problem context and fairness objectives. We provide a comprehensive mapping between each fairness measure and the belief system (i.e., worldview) that is encoded within its measurement of fairness. Lastly, we demonstrate the interpretability and efficacy of FINS in supporting the identification of real bias with case studies using AirBnB listings and voter records. 
    more » « less