Devices, systems, and methods for assessing and improving fairness of machine learning models with respect to entities that have protected characteristics

Abstract

A system for improving the fairness of a machine learning (ML) model includes an ML model and an electronic device. The ML model is configured to receive data regarding entities and provide a prediction for each entity that indicates whether it meets a particular criterion. The electronic device is configured to separate the predictions from the ML model into groups based on protected characteristics of entities that correspond to the predictions. The electronic device is configured to determine a fairness metric for the ML model based on the predictions and the grouping operation. After determining the fairness metric, the electronic device is configured to present the fairness metric to a user or trigger an alert if the fairness metric does not satisfy a fairness threshold. In this manner, the system can be used to improve the fairness of the ML model by highlighting biases thereof.

Description

DEVICES, SYSTEMS, AND METHODS FOR ASSESSING AND IMPROVING FAIRNESS OF MACHINE LEARNING MODELS WITH RESPECT TO ENTITIES THAT HAVE PROTECTED CHARACTERISTICSTECHNICAL FIELD

[0001] The present disclosure relates, generally, to machine learning (ML) and, more particularly, to ML models that handle data representing entities with protected characteristics.BACKGROUND

[0002] In recent years, the advent of sophisticated ML algorithms has revolutionized the way large volumes of data are processed, analyzed, and interpreted. These ML models, powered by artificial intelligence, have become indispensable tools across various industries, enabling the extraction of valuable insights from large and complex datasets.

[0003] One notable application of ML is in evaluating datasets to determine whether entities represented thereby meet particular criteria. A bank may leverage ML, for example, to determine whether to consider various applicants for a loan. Similarly, a company may use ML to narrow dow n the number of resumes before reaching out for interviews.

[0004] As powerful as ML models are, they are not immune to biases inherent in the data they are trained on. Biases present in the input data may be inadvertently learned by the ML algorithms, leading to skewed or unfair predictions. This is particularly true and especially problematic when the biases align with protected characteristics. For example, an ML model might inadvertently be trained such that it tends to provide positive predictions for male entities over female or non-male entities. The consequences of biased predictions are profound, potentially perpetuating existing social disparities and undermining principles of fairness and equity.SUMMARY

[0005] The present disclosure regards devices, systems, and methods that address some of the aforenoted problems, as well as other problems associated with ML and fairness. At a high level, the present disclosure teaches various means for assessing and improving the fairness of an ML model with respect to entities that have protected characteristics. These means may assist individuals, corporations, or other entities that wish to incorporate ML into theirworkflows (e.g., applicant assessment) while also avoiding the potential for discrimination that the ML might introduce. These means include characterizing and representing the fairness of an ML model such that someone unfamiliar with ML and / or fairness analysis can still understand and meaningfully correct for any potential biases of the ML model (e.g., by adjusting a workflow to account for the biases, by retraining or adjusting the ML model). These means also include monitoring the fairness of an ML model and reporting the fairness thereof via appropriate channels. Example embodiments of the advancements discussed herein include the following:

[0006] A system includes an ML model and an electronic device. The ML model is trained using training data that regards each example entity of a plurality of example entities and that indicates whether each example entity meets a particular criterion. Additionally, the ML model is configured to receive data regarding each entity of a plurality of entities. After receiving the data, the ML model then provides, for each entity, a respective prediction regarding whether the respective entity meets the particular criterion and a respective confidence score that corresponds to the respective prediction. The electronic device includes a display, a processor, and a non-transitory, computer-readable storage medium storing instructions that, when executed by the processor, cause the electronic device to perform various operations. The operations include providing the aforenoted data to the ML model. The operations also include receiving, from the ML model, (i) a plurality’ of predictions including the respective prediction for each entity and (ii) a plurality of confidence scores including the respective confidence score for each entity. Additionally, the operations include receiving, from a user input, (i) a plurality of protected characteristic indicators including a respective protected characteristic indicator for each entity that indicates whether the respective entity' has a particular protected characteristic and (ii) a plurality of targets including a respective target for each entity that indicates whether the respective entity meets the particular criterion. Further, the operations include designating, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions including the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions including the respective prediction for each entity that does not have the particular protected characteristic. Moreover, the operations include determining a fairness metric for the ML model based on (i) the first and second subsets of the plurality of predictions, (ii) the plurality of confidence scores, and (iii) the plurality of targets. The fairness metric indicates whether the ML model provides fair predictions with respect to entities that have the particular protectedcharacteristic and entities that do not have the particular protected characteristic. Furthermore, the operations include determining that the fairness metric does not satisfy a fairness threshold and then displaying a notice via the display. The notice indicates that the fairness metric does not satisfy the fairness threshold.

[0007] An electronic device includes a processor and a non-transitory, computer-readable storage medium storing instructions that, when executed by the processor, cause the electronic device to perform various operations. The operations include receiving, from an ML model, a plurality of predictions including a respective prediction for each entity of a plurality of entities, where the respective prediction regarding whether the respective entity meets a particular criterion. The operations also include receiving, from the ML model, a plurality of confidence scores including a respective confidence score for each entity, where the respective confidence score corresponds to the respective prediction. The ML model is configured to (i) receive data regarding each entity and (ii) provide, for each entity, the respective prediction and the respective confidence score. Additionally, the operations include receiving, from a user input, a plurality of protected characteristic indicators including a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic. Further, the operations include designating, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions including the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions including the respective prediction for each entity that does not have the particular protected characteristic. Moreover, the operations include determining a fairness metric for the ML model based on the first and second subsets of the plurality of predictions. The fairness metric indicates whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic. Furthermore, the operations include presenting the fairness metric via a display device associated with the electronic device.

[0008] A computer-implemented method includes procuring an ML model configured to (i) receive data regarding each entity’ of a plurality of entities and (ii) provide, for each entity, a respective prediction regarding whether the respective entity meets a particular criterion and a respective confidence score corresponding to the respective prediction. The method also includes receiving, from the ML model, (i) a plurality of predictions including the respective prediction for each entity of a plurality of entities and (ii) a plurality of confidence scores including the respective confidence score for each entity. Additionally, the method includesreceiving, from a user input, a plurality of protected characteristic indicators including a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic. Further, the method includes designating, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions including the respective prediction for each entity' that has the particular protected characteristic and (ii) a second subset of the plurality of predictions including the respective prediction for each entity that does not have the particular protected characteristic. Moreover, the method includes determining a fairness metric for the ML model based on the first and second subsets of the plurality' of predictions. The fairness metric indicates whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic. Furthermore, the method includes presenting the fairness metric via a display device.

[0009] It is understood that other configurations of the subject technology will become readily apparent to those skilled in the art from the following detailed description, wherein various configurations of the subject technology' are shown and described by way of illustration. As will be realized, the subject technology' is capable of other and different configurations and its several details are capable of modification in various other respects, all without departing from the scope of the subject technology. Accordingly, the figures and detailed description are to be regarded as illustrative in nature and not as restrictive.BRIEF DESCRIPTION OF THE FIGURES

[0010] For a better understanding of the various described embodiments, reference should be made to the Detailed Description below, in conjunction with the Figures. Like reference numerals refer to corresponding parts throughout the Figures and Description.

[0011] Figures 1 A and IB depict example systems for improving fairness of an ML model with respect to entities and protected characteristics, according to various aspects of the subject technology.

[0012] Figure 2 depicts an example process for improving fairness of an ML model with respect to entities and protected characteristics, according to various aspects of the subject technology.DETAILED DESCRIPTION

[0013] Reference will now be made to embodiments, examples of which are illustrated in the accompanying figures. In the following description, numerous specific details are set forth in order to provide an understanding of the various described embodiments. However, it will be apparent to one of ordinary skill in the art that the various described embodiments may be practiced without these specific details. In some instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.

[0014] Figures 1 A and IB depict example systems 100 and 150, respectively, for assessing and improving fairness of an ML model 104 with respect to entities and protected characteristics, according to various aspects of the subject technology. The first system 100, depicted in Figure 1 A, provides a high-level example of a fairness analysis device 102 that can be used to characterize, represent, and / or monitor the fairness of the ML model 104. The latter system 150, depicted in Figure IB, depicts the fairness analysis device 102 and the ML model 104 of the first system 100, along with an example fairness report 120. As discussed in more detail hereinbelow, in some embodiments, the fairness report 120 presents information regarding the fairness of the ML model 104 in a manner such that someone unfamiliar with ML and / or fairness analysis can comprehend the information contained therein and respond to it accordingly.

[0015] The fairness analysis device 102 is configured to allow a user to input information regarding the ML model 104 and / or predictions made thereby. The fairness analysis device 102 is also configured to process said information, as well as additional information pertinent to the ML model 104. Additionally, the fairness analysis device 102 is configured to present information regarding the fairness of the ML model 104 to the user (see, e.g., fairness report 120). As illustrated in Figure 1A, the fairness analysis device 102 includes an input / output (IO) interface 110, a processor 112, and a memory 114.

[0016] The IO interface 110 allows the user to input information into the fairness analysis device 102 and receive information therefrom. Accordingly, the IO interface 110 may include a keyboard, a mouse, a monitor display, a touchscreen, a cable connection interface (e.g., a USB port), a wireless connection interface (e.g., a Bluetooth antenna) and / or any other means for inputting information into the fairness analysis device 102 or presenting information to the user. The processor 112 of the fairness analysis device 102 is configured to execute instructionsstored in the memory 114, which may include volatile (e.g., random-access memory, cache memory) or non-volatile memory (e.g., read-only memory, flash memory, a magnetic storage device). In some embodiments, the memory includes a non-transitory, computer-readable storage medium storing instructions that, when executed by the processor 112, cause the fairness analysis device 102 to perform any of the various operations associated therewith and discussed herein.

[0017] In addition to the fairness analysis device 102, the example system 100 also includes the aforementioned ML model 104. As depicted in Figure 1 A, the ML model 104 may run on a server 106 or other electronic device external to the fairness analysis device 102. This is to illustrate that the subject technology does not require that the fairness analysis device 102 has direct access to the inner workings of the ML model 104 (e.g., an ML algorithm thereof). In some embodiments, the fairness analysis device 102 requires from the ML model 104 only its outputs (e.g., predictions and scores 118) in order to characterize, represent, and / or monitor the fairness thereof (e.g., the ML model 104 is a “black box” with respect to the fairness analysis device 102). In some embodiments, the ML model 104 is run on another device, such as the fairness analysis device 102 or another electronic device separate from the sen- er 106.

[0018] The ML model 104 is configured to receive data 116 regarding each entity' of a plurality of entities. This data 116 may be received from the fairness analysis device 102, but it may also be provided to the ML model 104 by another device separate from the fairness analysis device 102. Additionally, the ML model 104 is configured to provide, for each entity of the plurality of entities, a respective prediction regarding whether the respective entity7meets a particular criterion, as well as a respective confidence score that corresponds to the respective prediction (collectively, predictions and scores 118).

[0019] For example, the ML model 104 can be configured to provide a first pass on applications for a loan, ruling out applicants that are unlikely (e.g.. less than 20%, 30%, or 40% likely ) to be approved for the loan and therefore not worth reviewing further. In this example, the data 116 may regard applicants for the loan, including their respective credit score, credit history, annual income, debt-to-income ratio, and / or any other data relevant to eligibility for the loan (i.e., the criterion). After receiving this data 116, the ML model 104 can provide predictions and scores 118 for the applicants, indicating whether each applicant is likely (e.g., at least 60%, 70%, or 80% likely) to be approved for the loan and how confident the ML model 104 is in its prediction.

[0020] As another example, the ML model 104 can be configured to rule out a portion of resumes submitted for an open position at a company, saving the company time that it would otherwise need to spend reviewing the resumes and / or interviewing all of the applicants for the open position. In this example, the data 116 represents the applicants and includes information regarding the applicants’ qualifications for the job, as outlined in the applicants’ resumes and / or other application materials. For example, the data 116 may include applicants’ respective work experience, education (e.g.. highest education obtained), professional certifications, and / or other data relevant to the applicant’s eligibility for the open position. After receiving this data 116, the ML model 104 can provide predictions and scores 118 for the resumes regarding whether the applicant is worth interviewing for the position (e.g., meets all qualifications for the open position) and how confident the ML model 104 is regarding the same.

[0021] Configuring the ML model 104 to provide the predictions and scores 118 after receiving the data 116 involves training the ML model 104 with training data. At a high level, training the ML model 104 includes providing to the ML model 104 training data that regards each example entity of a plurality of example entities and that indicates whether each example entity meets the particular criterion. For example, training data for loan eligibility determinations may include data regarding example applicants (e.g.. including credit scores, annual salaries) and respective indicators of whether each applicant qualifies for a loan. As another example, training data for resume-related determinations may include data regarding example job applicants (e.g., including work experience, education) and respective indicators of whether each applicant ought to be interviewed for the job.

[0022] Unfortunately, biases inherent in the data used to train the ML model 104 may lead to downstream biases in the predictions and scores 118 it provides. This in turn may lead to unintentional discrimination with respect to entities (e.g., people) and protected characteristics. As used herein, protected characteristics refers to personal attributes (e.g., race, gender, age, disability, religion, sexual orientation) that are safeguarded against discrimination and unfair treatment in various social and legal contexts. By providing the predictions and scores 118 of the ML model 104 to the fairness analysis device 102, the fairness analysis device 102 can uncover said biases and inform or alert a user of the same.

[0023] The user may then use the information provided to them by the fairness analysis device 102 to adjust (e.g., retrain) the ML model 104 and / or adjust their response to the predictions and scores 118 it provides (e.g., by recalibrating internal policies with respect to review of applicants that have a particular protected characteristic). This may assist the user inensuring fair practices when said practices rely on ML. Moreover, this may assist as a mitigation control to guard against biased outcomes from the use of ML.

[0024] As illustrated in Figure IB, in system 150, the fairness analysis device 102 receives data from the ML model 104, including the predictions and / or confidence scores produced thereby (e g., predictions and scores 1 18). The data received by the fairness analysis device 1 2 need not be received directly from the ML model 104. In some embodiments, data from the ML model 104 is collected by a user and then provided by the user to the fairness analysis device 102 (e.g., in a JavaScript object notation (JSON) file and / or a comma-separated values (CSV) file).

[0025] The fairness analysis device 102 may also receive other data (e.g.. metadata) relevant to the fairness of the ML model 102, such as data regarding inputs provided to the ML model 104 (e.g., data 116), ground truth for the inputs provided to the ML model 104 (e.g., indicating what the ML model 104 should have predicted regarding whether said inputs meets the aforenoted criterion), information regarding protected characteristics associated with the inputs (e.g., indicating protected characteristics associated with the entities represented in the inputs), and / or training inputs used in training the ML model 104.

[0026] After receiving said predictions, scores, and / or other data from the ML model 104, the fairness analysis device 102 separates the predictions into (i) predictions that correspond to privileged persons 122 and (ii) predictions that correspond to unprivileged persons 124. (Alternatively, the predictions may be separated prior to being provided to the fairness analysis device 102.) As used herein, privileged and unprivileged is shorthand for the presence of a particular protected characteristic in an entity. For example, in certain contexts, the particular characteristic of male gender may be considered privileged whereas the characteristic of female or non-male gender may be considered unprivileged. Accordingly, the fairness analysis device 102 might separate predictions for male entities from predictions for female or nonmale entities. From there, the fairness analysis device 102 can determine whether the ML model 104 produces fair predictions with respect to the protected characteristic of gender (e.g., whether the ML model 104 produces different predictions with respect to the protected characteristic of gender).

[0027] This fairness determination may include calculating one or more metrics relating to the fairness of the ML model 104. These metrics can include (i) fairness metrics that represent the overall fairness of the ML model 104 and / or (ii) performance metrics that represent theperformance of the ML model 104 with respect to entities that have a particular protected characteristic (e.g., performance of the ML model 104 with respect to male entities and / or female or non-male entities).

[0028] Following the fairness determination, the fairness and / or performance metrics can be presented to a user in a fairness report 120. The fairness report 120 may include visual representations of the metrics, such as in visual section 126. These visual representations may assist laypeople (e.g., bank employees and / or company recruiting personnel) in understanding biases of the ML model 104. For example, the visual section 126 includes a speedometer dial 128 with aneedle 130 suggesting that the ML model 104 bears a moderate bias towards a privileged (“P”) group. In addition to a speedometer-like representation, other example visual representations include bar graphs (e.g., representing a decile analysis, a statistical significance test, and / or group size and corresponding outcome rate ), line graphs (e.g., representing a receiver operating characteristic curve, a precision recall curve, and / or a calibration curve), confusion matrices (e.g., an N-by-N matrix for summarizing the performance of an ML model, where N is the number of target classes), and / or scatter plots (e.g., apredicted-vs-actual scatter plot and / or a predicted-vs-residual scatter plot). These visual representations can draw out differential treatment by the ML model 104 between the groups (e.g., allowing for a root cause analysis).

[0029] The fairness report 120 may also include numerical representations of the fairness and / or performance metrics, such as the numerical section 132. These numerical representations may include brief explanations of the metrics represented therein to further assist laypeople in interpreting the fairness report 120 and responding accordingly. Thus, the fairness report 120 may include different sections, with each section providing a different piece of information that may help in identifying biases of the ML model 104.

[0030] The various fairness and performance metrics determined for the ML model 104 may depend in part on the type of the ML model 104. For example, if the ML model 104 is a binary classification model, the fairness analysis device 102 may determine fairness metrics that include (i) an average odds difference metric, (ii) a disparate impact metric, (iii) a statistical parity difference metric, (iv) an equal opportunity difference metric, (v) a Theil index metric, (vi) a true negative rate ratio, (vii) a false negative rate ratio, (viii) a false discovery rate ratio, (ix) a false omission rate ratio, (x) a true positive rate ratio, (xi) a false positive rate ratio, (xii) a precision ratio, (xiii) a predicted positive rate ratio, (xiv) a predicted prevalence rate ratio, and / or (xv) a negative predicted value ratio. For a binary classification model, the fairnessanalysis device 102 may also determine performance metrics that include (i) an accuracy metric, (ii) an area-under-curve (AUC) metric, (iii) a log loss metric, (iv) a precision metric, (v) a recall metric, and / or (vi) an F 1 score.

[0031] However, if the ML model 104 is a regression model, the fairness analysis device 102 may determine fairness metrics that include (i) a demographic parity difference metric, (ii) a direct density measure of independence, (iii) a direct density measure of separation, and / or (iv) a direct density measure of sufficiency. For a regression model, the fairness analysis device 102 may also determine performance metrics that include (i) a mean squared error metric, (ii) a mean absolute error metric, (iii) a root mean square error metric, (iv) a median absolute deviation metric, (v) a mean absolute percentage error metric, (vi) a mean prediction metric, (vii) a mean overprediction metric, (viii) a mean underprediction metric, (ix) a coefficient of determination, (x) a Pearson correlation metric, and / or (xi) a Spearman correlation metric.

[0032] For reference, these fairness and performance metrics are summarized below. Each summary includes a definition and formula for the respective metric, as well as a description of how the metric might bear on the fairness of the ML model 104. In some embodiments, portions of these definitions, formulas, and / or descriptions are included in the fairness report 120 to assist the user in interpreting the fairness report 120.

[0033] Average Odds Difference. This is a measure of the mean of (i) the difference in the false positive rates of the unprivileged and privileged groups and (i) the difference in the true positive rates of the unprivileged and privileged groups. It stems from the statistical notion of fairness called equalized odds, which states that a classifier must have equal true positive and false positive rates for all groups. For fair outcomes, the false positive rates and the true positive rates for the unprivileged and privileged groups should be same. Hence, a value of zero indicates exact fairness. A non-zero value indicates that the false positive rates and true positive rates are different between the two groups. Moreover, a positive value indicates that the privileged group has a lower false positive rate and / or true positive rate. Likewise, a negative indicates that the unprivileged group has a lower false positive rate and / or true positive rate. Average odds difference can be calculated as follows:(FPRunpriv ~ FPRpriv) + (TPRunpriv FP Rpriv)2 whereTP FPTPR = - TP+FN and FPR = - FP+TN and where FPR denotes false positive rate, TPR denotes true positive rate, TP denotes true positive predictions, TN denotes true negative predictions, FP denotes false positive predictions, and FN denotes false negative predictions.

[0034] Disparate Impact. This is a measure of the difference in positive predictions between the unprivileged group and the privileged group. It is calculated as the ratio of the rate at which the unprivileged group receives positive predictions to the rate at which the privileged group receives positive predictions. This metric does not take ground truth into consideration and only depends on the predictions of the ML model 104. For fairness, the positive outcome rates between unprivileged and privileged group should be same. A value of one indicates that the ML model 104 is fair. A value greater than one indicates lower positive outcome rates for the unprivileged group, which suggests that the ML model 104 is biased against the unprivileged group. A value less than one indicates lower positive outcome rates for the privileged group, suggesting a bias against the privileged group. Disparate impact can be calculated as follows:where TP denotes true positive predictions, FP denotes false positive predictions, and N denotes the total number of predictions for a particular group.

[0035] Statistical Parity Difference. This is a measure of the difference in positive outcomes between the unprivileged group and the privileged group. It is calculated as the difference between the rate at which the unprivileged group receives positive predictions and the rate at which the privileged group receives positive predictions. Accordingly, it is used to check for equalized outcomes across the privileged and unprivileged groups. This metric does not take ground truth into consideration and only depends on the predictions from the ML model 104. For fairness, the positive outcome rates between unprivileged and privileged group should be same (e.g. people across groups should have the same chance of being approved for a loan or invited for an interview). Hence, a value of zero indicates fairness. A negative value indicates lower positive outcome rates for the unprivileged group, suggesting a bias against the unprivileged group. A positive value indicates lower positive outcome rates for the privileged group, suggesting a bias against the privileged group. Statistical parity difference can be calculated as follows:where TP denotes true positive predictions, FP denotes false positive predictions, and N denotes the total number of predictions for a particular group.

[0036] Equal Opportunity Difference. This is a measure of the difference between the true positive rate of the unprivileged group and that of the privileged group. For fairness, the true positive between unprivileged and privileged groups should be same. Hence, a value of zero indicates fairness. A negative value means a lower TPR for the unprivileged group as compared to the privileged group. And a positive value means a lower TPR for the privileged group - indicating a bias against the privileged group. Equal opportunity difference can be calculated as follows:T1 rP ^Runpriv — T1 rP ^Rpriv whereTP TPR~ TP + FN and where TPR denotes true positive rate, TP denotes true positive predictions, and FA denotes false negative predictions.

[0037] Theil Index. This is a special case of a family of inequality indices called generalized entropy indices. It is an indicator of the overall level of disproportionate predictions in the entire population. The value of Theil index ranges between zero and infinity, where zero represents an equal distribution and positive values represent increasing amounts of disproportion. An equal distribution of predictions is indicated by a value close to zero. The Theil index value remains the same for a particular dataset irrespective of the protected characteristic in question as it indicates the overall noise present in the predictions. It can be calculated as follows:whereTpred - ytrue for the ithinstanceand where ypred denotes the predicted value for the respective prediction, ytrUe denotes the ground truth for the respective prediction, and N denotes the total number of predictions for a particular group.

[0038] True Negative Rate Ratio. This is the ratio of the true negative rate between the unprivileged and privileged groups, where true negative rate indicates the proportion of true negatives correctly classified as negative by the ML model 104. For fairness, the true negative rate ratio between the unprivileged and privileged groups should be the same. Hence, a value of one indicates fairness. A positive value means a higher true negative rate for the unprivileged group as compared to the privileged group. A negative value means a higher true negative rate for the privileged group as compared to the unprivileged group. True negative rate ratio can be calculated as follows:TNRunprivTNRprivwhereTNTNRTN + FP and where TNR denotes true negative rate, TN denotes true negative predictions, and FP denotes false positive predictions.

[0039] False Negative Rate Ratio. This is the ratio of the false negative rate between the unprivileged and privileged groups, where false negative rate indicates the proportion of true positives incorrectly classified as negative by the ML model 104. For fairness, the false negative rate ratio between the unprivileged and privileged groups should be the same. A value of one indicates exact fairness. A value greater than one means a higher false negative rate for the unprivileged group as compared to the privileged group. A value below one means a higher false negative rate for the privileged group as compared to the unprivileged group. False negative rate can be calculated as follows:1F7NVR'-unpnvFNRprivwhereFNFNRFN + TPand where FNR denotes false negative rate, FN denotes false negative predictions, and TP denotes true positive predictions.

[0040] False Discovery Rate Ratio. This is the ratio of the false discovery rate between the unprivileged and privileged groups, where false discovery rate indicates the proportion of true negatives incorrectly classified as positive by the ML model 104. For fairness, the false discovery' rate ratio between the unprivileged and privileged groups should be same. Hence, a value of one indicates exact fairness. A value greater than one indicates a higher false discoveryrate for the unprivileged group as compared to the privileged group. And a value less than one indicates a higher false discovery' rate for the privileged group as compared to the unprivileged group. False discovery' rate ratio can be calculated as follows:FDRunprivFDRprivwhereFPFDRFP + TP and where FDR denotes false discovery rate, FP denotes false positive predictions, and TP denotes true positive predictions.

[0041] False Omission Rate Ratio. This is the ratio of the false omission rate between the unprivileged and privileged groups, where false omission rate indicates the proportion of negative predictions incorrectly classified as negative by the ML model 104. For fairness, the false omission rate ratio between the unprivileged and privileged groups should be same. Hence, a value of one indicates exact fairness. A value greater than one indicates a higher false omission rate for the unprivileged group as compared to the privileged group. A value less than one indicates a higher false omission rate for the privileged group as compared to the unprivileged group. False omission rate ratio can be calculated as follows:FORunpr[VFORprlvwhereFNFORFN + TN and where FOR denotes false omission rate, FN denotes false negative predictions, and TN denotes true negative predictions.

[0042] True Positive Rate Ratio. This is the ratio of the true positive rate between the unprivileged and privileged groups, where true positive rate is the proportion of true positives in the dataset correctly classified as positives by the ML model 104. It is achieved if the sensitivities in the subgroups are close to each other. For fairness, the true positive rate ratio between the unprivileged and privileged groups should be same. Hence, a value of one indicates exact fairness. A value greater than one indicates a higher true positive rate for the unprivileged group as compared to the privileged group. A value less than one indicates a true positive rate for the privileged group as compared to the unprivileged group. True positive rate ratio can be calculated as follows:T P R-unpriv TPRpriv whereTPTPR~ TP + FN and where TPR denotes true positive rate, TP denotes true positive predictions, and FN denotes false negative predictions.

[0043] False Positive Rate Ratio. This is the ratio of the false positive rate between the unprivileged and privileged groups, where false positive rate indicates the proportion of true negatives in the dataset incorrectly classified as positive by the ML model 104. False positive rate parity is achieved if the false positive rates in the subgroups are close to each other. False positive rate is important when the cost of incorrectly identifying a positive is high (e.g., due to additional work or expense). For fairness, the false positive rate between the unprivileged and privileged groups should be same. Hence, a value of one indicates exact fairness. A value greater than one indicates a higher false positive rate for the unprivileged group as compared to the privileged group. A value less than one indicates a higher false positive rate for the privileged group as compared to the unprivileged group. False positive rate ratio can be calculated as follows:F Runpriv 1 F1Pi-priv whereFPFPRFP + TNand where FPR denotes false positive rate, FP denotes false positive prediction, and TN denotes true negative prediction.

[0044] Precision Ratio. This is the ratio of the precision between the unprivileged and privileged groups, where precision indicates the proportion of predicted positives that are actually positive. It is achieved if the precision in the subgroups are close to each other. For fairness, the precisions between the unprivileged and privileged groups should be same. Hence, a value of one indicates exact fairness. A value greater than one indicates a higher precision for the unprivileged group as compared to the privileged group. A value less than one indicates a higher precision for the privileged group as compared to the unprivileged group. Precision ratio can be calculated as follows: precisionunprivprecisionprivwhereTP precision =Tp + Fpand where TP denotes true positive predictions, and FP denotes false positive predictions.

[0045] Predicted Positive Rate Ratio. This is the ratio of the predicted positive rate between the unprivileged and privileged groups, where predicted positive rate is the ratio of positive predictions by the ML model 104 for a group to the total number of predictions for both the privileged and unprivileged groups. As with some of the other fairness metrics, predicted positive rate ratio does not rely on ground truth and depends only on the ML model 104 predictions. For fairness, the predicted positive rate ratio between the unprivileged and privileged groups should be the same. Thus, a predicted positive rate ratio of one indicates exact fairness. A value greater than one indicates a higher predicted positive rate for the unprivileged group as compared to privileged group, and a value less than one indicates a higher predicted positive rate for the privileged group. Predicted positive rate ratio can be calculated as follows:PP unpriv PPRpriv where nppand where PPR denotes predicted positive rate, TP denotes true positive predictions, FP denotes false positive predictions, and N denotes the total number of predictions for a particular group.

[0046] Predicted Prevalence Rate Ratio. This is the ratio of the predicted prevalence rate predicted prevalence rate between the unprivileged and privileged groups, where predicted prevalence rate indicates the fraction of positive predictions by the classifier. This metric does not depend on ground truth. For fairness, the predicted prevalence rate ratio between the unprivileged and privileged groups should be the same. A value of one indicates exact fairness, a value greater than one indicates a higher predicted prevalence rate ratio for the unprivileged group as compared to the privileged group, and a value less than one indicates a higher predicted prevalence rate ratio for the privileged group. Predicted prevalence rate ratio can be calculated as follows:PPrevRunprivPPrevRprivwhereand where PPrevR denotes predicted prevalence rate, TP denotes true positive predictions, FP denotes false positive predictions, and A denotes the total number of predictions for a particular group.

[0047] Negative Predicted Value Ratio. This is the ratio of the negative predicted value between the unprivileged and privileged groups, where negative predicted value indicates the proportion of predicted negatives that are actually negative. Negative predicted value parity is achieved if the negative predictive values in the subgroups are equal to each other. For fairness, the negative predicted value ratio between the unprivileged and privileged groups should be the same. Accordingly, a value of one indicates exact fairness. A value greater than one indicates a higher negative predicted value for the unprivileged group, and a value less than one means a higher negative predicted value for the privileged group. Negative predicted value ratio can be calculated as follows:NPVunpriv NPVprivwhereTNNPV =TN + FN and where NPV denotes negative predicted value, TN denotes true negative predictions, and FN denotes false negative predictions.

[0048] Demographic Parity Difference. This is a measure of the difference between the mean predicted outcome for two different demographic groups (e.g., male and female entities). For a fair outcome, the mean prediction for the unprivileged and privileged groups should be same. Hence, a value of zero indicates fairness, and a non-zero value indicates that mean predictions are different between the two groups. Demographic parity' difference can be calculated as follows:where DPD denotes demographic parity difference, n denotes the number of predictions for the privileged group, m denotes the number of predictions for the unprivileged group, and Y denotes predictions for a group.

[0049] Direct Density Measures. Direct density measures, including independence, separation, and sufficiency (see below), are used to assess whether the predictions of the ML model 104 are fair across different demographic groups. These measures are designed to evaluate the extent to which the predictions of the ML model 104 are based on the protected characteristic (e g., race, gender) of the entities being evaluated.

[0050] Direct Density Measure of Independence. This measures the extent to which the protected characteristic is independent of the predictions of the ML model 104. If the ML model 104 is fair, the probability of a certain outcome should be the same for all entities, regardless of any protected characteristics associated therewith. For fair outcomes, the independence measure should be between 0.8 and 1.25. A value greater than 1.25 indicates a bias against the unprivileged group. Direct density' measure of independence can be calculated as follows:S Awhere ru denotes independence ratio, A denotes protected data in privileged and unprivileged groups, and S' denotes predicted values.

[0051] Direct Density Measure of Separation. This measures the extent to which the predictions of the ML model 104 are based on the protected characteristic. If the ML model 104 is fair, the protected characteristic should not be a significant predictor of the outcome. For fair outcomes, the value should be between 0.8 and 1.25. A value greater than 1.25 indicates a bias against the unprivileged group. Direct density measure of separation can be calculated as follows:S 1 A | Y_ P(S\A = 1, Y) TSEP =P(S\A = o, r) where rsepdenotes separation ratio, A denotes protected data in privileged and unprivileged groups, S denotes predicted values, and Y denotes observed values.

[0052] Direct Density Measure of Sufficiency. This measures the extent to which the protected characteristic is sufficient to explain the predictions of the ML model 104. If the ML model 104 is fair, the protected characteristic should not be a significant predictor of the outcome after controlling for other relevant variables. For fair outcomes, the value should be between 0.8 and 1.25. A value greater than 1.25 indicates a bias against the unprivileged group. Direct density measure of sufficiency can be calculated as follows:Y 1 A | S_ P(T|4 = 1,S) Vsuf =P(Y\A = 0,5) where rsuf denotes sufficiency ratio. A denotes protected data in privileged and unprivileged groups, S denotes predicted values, and Y denotes observed values.

[0053] Accuracy. This is the ratio of correct predictions to the total number of predictions made by the ML model 104. The highest possible value for accuracy is one and the lowest possible value is zero. A value higher for privileged group means that the accuracy is better for privileged group as compared to unprivileged group. A value higher for the unprivileged group means that the accuracy is better for unprivileged group as compared to privileged group. A similar value for both groups means that the ML model 104 has similar accuracy for privilegedand unprivileged group, suggesting that the ML model 104 is fair. Accuracy can be calculated as follows:and where A denotes accuracy, TP denotes the number of true positive predictions, TN denotes the number of true negative predictions, and N denotes the total number of predictions.

[0054] AUC. This is a measure of the area under the receiver-operating-characteristic (ROC) curve. The ROC curve is a graphical plot that illustrates the diagnostic ability of the ML model 104 as its discrimination threshold is varied. An ROC cune plots true positive rate versus false positive rate at different classification thresholds. AUC provides an aggregate measure of performance across all possible classification thresholds. One way of interpreting AUC is as the probability’ that the ML model 104 ranks a random positive example more highly than a random negative example. The higher the area under the curve, the better the performance of the ML model 104. AUC ranges in value from 0 to L A value higher for the privileged group means that the AUC score is better for privileged group as compared to the unprivileged group, and vice versa. A similar value for both groups means that the ML model 104 has similar AUC scores for the privileged and unprivileged groups, suggesting that the ML model 104 is fair.

[0055] Log Loss. This is the negative average of the logarithm of corrected predicted probabilities for each instance. It is a measure of the performance of a classification model whose output is a probability. Log loss increases as the predicted probability diverges from the actual label. Thus, log loss is indicative of how close the prediction probability is to the corresponding actual value. Its value can be between 0 and infinity. The more the predicted probability diverges from the actual value, the higher is log loss value. For any given problem, a lower log loss value means better predictions. A perfect model would have a log loss of 0. Log loss is a good metric for comparing two ML models, provided both of the ML models are applied to the same distribution of dataset. A similar log loss value for the privileged and unprivileged groups means that the ML model 104 has similar log-loss values for each group, suggesting that the ML model 104 is fair. Log loss can be calculated as follows:where N denotes the total number of predictions. v;denotes the actual class of the instance, p(y) is the probability of the instance belonging to the class 1 from the classifier, and (l-pfyi)) is the probability of the instance belonging to the class 0 from the classifier.

[0056] Precision. This is the ratio of correct positive predictions to the total number of positive predictions made by the ML model 104. The highest possible value of precision is one and the lowest possible value is zero. A higher value for the privileged group means that the precision score is better for the privileged group as compared to the unprivileged group, and vice versa. A similar value for both groups means that the ML model 104 has similar precision for the privileged and unprivileged groups, suggesting that the ML model 104 is fair. Precision can be calculated as follows:TP TP + FP where TP denotes the true positive predictions and FP denotes the false positive predictions.

[0057] Recall. This is the ratio of correct positive predictions to the overall number of positive instances in the dataset. The highest possible value of recall is one, and the lowest possible value is zero. A higher value for the privileged group means that the recall score is better for the privileged group as compared to the unprivileged group, and vice versa. A similar value for both groups means that the ML model 104 has similar recall for the privileged and unprivileged groups, suggesting that the ML model 104 is fair. Recall can be calculated as follows:TP TP + FN where TP denotes the true positive predictions and FN denotes the false negative predictions.

[0058] Fl Score. This is the harmonic mean of precision and recall. The highest possible value of the Fl score is one, indicating perfect precision and recall, and the lowest possible value is 0, if either the precision or the recall is zero. A higher value for the privileged group means that the Fl score is better for the privileged group as compared to unprivileged group, and vice versa. A similar value for both groups means that the ML model 104 has similar Fl score for the privileged and unprivileged groups, suggesting that the ML model 104 is fair. Fl score can be calculated as follows:

[0059] Mean Squared Error. This is the average of the squared differences between the predicted and actual values (i.e.. ground truth). The square of the difference is used to ensure that both positive and negative differences contribute to the overall error. Mean squared error can be used to compare the performance of different ML models, with lower mean squared error indicating better performance. Mean squared error can be calculated as follows:where MSE denotes mean squared error, n denotes the number of data points, Y denotes observed values, and T, denotes predicted values.

[0060] Mean Absolute Error. This is the average of the absolute differences between the predicted and actual values. Unlike mean squared error, mean absolute error does not square the differences, so it gives equal weight to both positive and negative errors. Mean absolute error can also be used to compare the performance of different models, with lower mean square error values indicating better performance. Mean absolute error can be calculated as follows:where MAE denotes mean absolute error, y denotes predictions, x, denotes true values (i.e., ground truth), and n denotes the total number of data points.

[0061] Root Mean Square Error. This is the square root of the average of the squared differences between the predicted and actual values. Like mean squared error, root mean square error gives more weight to larger errors, but taking the square root gives the metric the same units as the predicted variable. Root mean square error can be used to compare the performance of different models, with lower root mean square error indicating better performance. Root mean square error can be calculated as follows:where RMSE denotes root mean square error, N denotes the number of data points, x;denotes actual observations time series, and;denotes estimated time series.

[0062] Median Absolute Deviation. This is a measure of the variability or dispersion of the dataset and is often used as a robust alternative to standard deviation, especially when dealing with datasets that contain outliers or extreme values. It is calculated by taking the median ofall absolute differences between the target and the prediction. Median absolute deviation can be used to measure the variability of a dataset and provide a measure of how spread out a set of data is. It can also be used in combination with other measures, such as the median metric, to provide a robust measure of central tendency and variability. Median absolute deviation can be calculated as follows:where MAD denotes median absolute deviation, Xi denotes each value, and X denotes average value.

[0063] Mean Absolute Percentage Error. This is the absolute difference between the predicted and actual values, divided by the actual value, and then averaged across all observations in the dataset. Mean absolute percentage error provides a measure of the accuracy of a regression-type ML model, with lower values indicating better performance. Mean absolute percentage error can be calculated as follows:where MADE denotes mean absolute percentage error, n denotes the total number of predictions, Atdenotes actual values, and Ftdenotes predicted values.

[0064] Mean Prediction. This is the average of all predictions made by the ML model 104. Mean predication can be calculated as follows:where MP denotes mean prediction, n denotes the number of predictions, and Yt denotes predicted values.

[0065] Mean Overprediction. This is a statistical measure that indicates the average amount by which a prediction exceeds the actual value. It can be used to assess the performance of the ML model 104 and identify areas where the ML model 104 may be consistently overestimating the actual values. Additionally, it can be used in combination with other measures, such as mean absolute error and mean squared error, to provide a comprehensive evaluation of the performance of the ML model 104. Mean overprediction can be calculated as follows:where MOP denotes mean overprediction, n denotes total number of predictions. Yi denotes actual values, and denotes predicted f, values.

[0066] Mean Underprediction. This is a statistical measure that indicates the average amount by which a prediction falls short of the actual value. It can be used to assess the performance of the ML model 104 and identify areas where the ML model 104 may be consistently underestimating the actual values. Additionally, it can be used in combination with other measures, such as mean absolute error and mean squared error, to provide a comprehensive evaluation of the performance of the ML model. Mean underprediction can be calculated as follows:where MUP denotes mean underprediction, n denotes total number of predictions. T;denotes actual values, and denotes predicted Y values.

[0067] Coefficient of Determination. This is a statistical measure that indicates how well the ML model 104 fits the observed data. Specifically, coefficient of determination measures the proportion of the variance in the dependent variable that is explained by the independent variable(s) in the ML model 104. It is a value between 0 and 1, with higher values indicating a better fit. A value of 1 indicates that the ML model 104 explains all of the variance, while a value of 0 indicates that the ML model 104 does not explain any of the variance. Coefficient of determination can be calculated as follows:, RSSR~1~ TSS where R2denotes coefficient of determination, RSS denotes the sum of squares of residuals, and TSS denotes the total sum of squares.

[0068] Pearson Correlation. This is a statistical measure that indicates the strength and direction of the linear relationship between two variables. It can be used to determine the strength of the relationship between the dependent and independent variables. However, it assumes that the relationship between the two variables is linear, and that the data follows a normal distribution. Pearson correlation coefficient ranges from -1 to 1. A value of -1 indicatesa perfect negative correlation, meaning that as one variable increases, the other decreases. A value of 1 indicates a perfect positive correlation, meaning that as one variable increases, the other increases as well. A value of 0 indicates no linear correlation between the two variables. Pearson correlation can be calculated as follows:where r denotes the correlation coefficient, x, denotes the values of the x-variable in a sample, x denotes the mean of the values of the x-variable, y, denotes the values of the y-variable in a sample, andy denotes the mean of the values of the y-variable.

[0069] Spearman Correlation. This is a statistical measure that indicates the strength and direction of the monotonic relationship between two variables. It is a non-parametric measure of correlation, meaning that it does not assume that the relationship between the two variables is linear or that the data follows a normal distribution. It is useful when outliers or extreme values may be present in the data. Specifically, Spearman’s correlation measures the degree of association between two variables by ranking the observations for each variable and then calculating the Pearson correlation coefficient for the ranks. It ranges from -1 to 1, with a value of -1 indicating a perfect negative monotonic relationship, a value of 1 indicating a perfect positive monotonic relationship, and a value of 0 indicating no monotonic relationship. Spearman correlation can be calculated as follows:where p denotes the Spearman correlation, di denotes the difference between the two ranks of each observation, and n denotes the number of observations.

[0070] Figure 2 depicts an example process 200 for assessing and improving fairness of an ML model (e.g., ML model 104) with respect to entities and protected characteristics, according to various aspects of the subject technology. One or more blocks of the process 200 may be implemented, for example, by one or more computing devices, such as the fairness analysis device 102 of Figures 1A and IB and / or the server 106 of Figure 1A.

[0071] In some embodiments, one or more of the blocks may be implemented based on one or more ML algorithms. In some embodiments, one or more of the blocks may be implemented apart from other blocks, and by one or more different processors or devices. Further, for explanatory purposes, the blocks of the process 200 are described as occurring in serial (i.e.,linearly). However, some of the blocks of the process 200 may occur in parallel (i.e., simultaneously). Additionally, the blocks of the process 200 need not be performed in the order shown and one or more of the blocks of the process 200 need not be performed.

[0072] In the depicted example, a processor of an electronic device (e.g.. fairness analysis device 102) receives (202) a plurality of predictions and a plurality of confidence scores (e.g., predictions and scores 118) from an ML model (e.g., ML model 104). The plurality of predictions includes a respective prediction for each entity of a plurality' of entities, and the plurality of confidence scores includes a respective confidence score for each entity. Moreover, the ML model is configured to receive data (e.g., data 116) regarding each entity of the plurality of entities and provide, for each entity, a respective prediction regarding whether the respective entity meets a particular criterion (e.g., loan eligibility', job qualification) and a respective confidence score that corresponds to the respective prediction.

[0073] In some embodiments, the ML model is trained using training data that regards each example entity of a plurality of example entities and that indicates whether each example entity meets the particular criterion. For example, as discussed above, the training data may include data regarding example people and respective indications regarding whether each person is eligible for a loan or meets the requirements for a job position. Accordingly, as an example, the ML model can be trained to determine whether people are eligible for the loan or meet the requirements for the job position.

[0074] In addition to the pluralities of predictions and confidence scores, the processor also receives (204) a plurality of protected characteristic indicators from a user input (e.g., in a JSON file and / or a CSV file). The plurality of protected characteristic indicators includes a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic. For example, if the entities are people and the protected characteristic at issue is nationality, then the plurality of protected characteristic indicators may indicate the nationality of each person for which the ML model 104 made a prediction.

[0075] After receiving the pluralities of predictions, confidence scores, and protected characteristic indicators, the processor designates (206) first and second subsets of the plurality of predictions that correspond, respectively, to entities that do and do not have a particular protected characteristic (e.g., foreign nationality, domestic nationality). This designation may involve separating the plurality of predictions into a first subset of predictions for entities withthe protected characteristic (e.g., foreign nationality) and a second subset of predictions for entities without the protected characteristic (e.g., domestic nationality).

[0076] In some embodiments, the processor receives the aforenoted first and second subsets of the plurality of predictions (or an indication of which portions of the plurality of predictions correspond to which entities and / or which protected characteristics). In these embodiments, there may be no need for the processor to receive the non-separated plurality of protected characteristic indicators or for the processor to designate the first and second subsets of the plurality of predictions.

[0077] Based on the first and second subsets of the plurality of predictions, the processor determines (208) (e.g.. calculates, computes) a fairness metric for the ML model or, in some embodiments, a performance metric relating thereto. As discussed above, the fairness metric may indicate whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic, and the performance metric may regard the performance of the ML model with respect to the first or second subset of the plurality of predictions.

[0078] In some embodiments, the fairness or performance metric determination may also be based on a plurality of targets (e.g.. ground truths). The plurality of targets may include, for example, a respective target for each entity, with the respective target indicating whether the respective entity meets the particular criterion. In these embodiments, the processor may receive this plurality of targets from the aforenoted user input.

[0079] In some embodiments, the processor may receive an indicator of a ty pe of the ML model prior to determining the fairness metric or the performance metric, where the model type indicator indicates whether the ML model is a binary classification model or a regression model. If the ML model is a binary classification model, then the processor may determine any of the metrics specific to binary classification models, discussed above. Likewise, if the ML model is a regression model, then the processor may determine any of the metrics specific to regression models, also discussed above.

[0080] After determining the fairness or performance metric, the processor presents (210) the fairness or performance metric via a display device, such as a display of the fairness analysis device 102 or a display connected thereto (e.g., via network 108). Presenting the fairness or performance metric may include generating a graphic representation of the fairness metric (e.g., fairness report 120) and displaying it to a user. In some embodiments, presenting the fairnessor performance metric includes transmitting an indication of the fairness or performance metric to the user (e.g., to a mobile device or a personal computer of the user).

[0081] Additionally, or alternatively, after determining the fairness or performance metric, the processor may compare the fairness or performance metric to a first fairness threshold. Or the processor may compare the performance metric and another performance metric (e.g., a difference between the two performance metrics) to another fairness threshold. In the event that either of the fairness thresholds are not met, the processor may trigger an alert indicating, for example, that the fairness metric does not satisfy the first fairness threshold. This alert may include a notification, an email, or any other means whereby the disparity between metric and threshold might be communicated. Moreover, this alert may include information for assisting people unfamiliar with ML and / or fairness analysis in understanding the severity of the fairness issue and how to go about correcting the issue or accounting for it.

[0082] In some embodiments, the processor adjusts the ML model based on the fairness metric or the performance metric. For example, the processor may adjust internal weights of the ML model to improve the fairness or the performance thereof. As another example, the processor may retrain the ML model by providing additional training data to the ML model. The additional training data, for example, may not include biases inherent in training data originally used to train the ML model.

[0083] In some embodiments, the performance metric determination is based on the aforenoted plurality of targets, as well as the first and second subsets of the plurality of predictions. Further, in some of these embodiments, the processor also determines a second performance metric, w here the first performance metric corresponds to the performance of the ML model with respect to the first subset of the plurality' of predictions and the second performance metric corresponds to the second subset of the same. These two performance metrics may be of the same type. For example, the processor may determine the precision of the ML model with respect to privileged and nonprivileged entities. In this manner, a difference between the two metrics may highlight a potential bias in the ML model.

[0084] In some embodiments, the plurality of entities is a plurality of people, where each respective entity is a respective person. In some of these embodiments, the protected characteristic is a particular type of protected characteristic, such as age, disability status, gender, marital status, maternity’ status, national origin, race, religion, sexual orientation, or socioeconomic status.

[0085] Illustration of Subject Technology as Clauses:

[0086] Various examples of aspects of the present disclosure are described as numbered clauses below. These are provided as examples and are not intended limit the subject technology. Identifications of the figures and reference numbers are provided below merely as examples and for illustrative purposes, and the clauses are not limited by these identifications.

[0087] Clause 1. A system for improving fairness of an ML model with respect to entities and protected characteristics, the system comprising: an ML model (i) trained using training data that regards each example entity7of a plurality of example entities and that indicates whether each example entity meets a particular criterion and (ii) configured to receive data regarding each entity of a plurality of entities and provide, for each entity, a respective prediction regarding whether the respective entity meets the particular criterion and a respective confidence score corresponding to the respective prediction; and an electronic device comprising a display, a processor, and a non-transitory, computer-readable storage medium storing instructions that, yvhen executed by the processor, cause the electronic device to: provide the data to the ML model; receive, from the ML model, (i) a plurality of predictions comprising the respective prediction for each entity and (ii) a plurality of confidence scores comprising the respective confidence score for each entity; receive, from a user input, (i) a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity that indicates whether the respective entity7has a particular protected characteristic and (ii) a plurality of targets comprising a respective target for each entity that indicates whether the respective entity7meets the particular criterion; designate, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions comprising the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic; determine a fairness metric for the ML model based on (i) the first and second subsets of the plurality7of predictions, (ii) the plurality of confidence scores, and (iii) the plurality of targets, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; determine that the fairness metric does not satisfy a fairness threshold; and display a notice via the display responsive to determining that the fairness metric does not satisfy the fairness threshold, the notice indicating that the fairness metric does not satisfy the fairness threshold.

[0088] Clause 2. An electronic device for determining fairness of an ML model with respect to entities that have a protected characteristic, the electronic device comprising a processor and a non-transitory. computer-readable storage medium storing instructions that, when executed by the processor, cause the electronic device to: receive, from an ML model, (i) a plurality of predictions comprising a respective prediction for each entity7of a plurality7of entities, the respective prediction regarding whether the respective entity meets a particular criterion and (ii) a plurality of confidence scores comprising a respective confidence score for each entity, the respective confidence score corresponding to the respective prediction, the ML model being configured to (i) receive data regarding each entity and (ii) provide, for each entity, the respective prediction and the respective confidence score; receive, from a user input, a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity that indicates whether the respective entity7has a particular protected characteristic; designate, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality7of predictions comprising the respective prediction for each entity7that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic; determine a fairness metric for the ML model based on the first and second subsets of the plurality of predictions, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; and present the fairness metric via a display device associated with the electronic device.

[0089] Clause 3. The system of Clause 2, wherein the ML model is trained using training data (i) that regards each example entity of a plurality of example entities and (ii) that indicates whether each example entity7meets the particular criterion.

[0090] Clause 4. The system of either Clause 2 or Clause 3, wherein determining the fairness metric is further based on a plurality7of targets comprising a respective target for each entity that indicates whether the respective entity meets the particular criterion, and the instructions further cause the electronic device to receive the plurality of targets from the user input.

[0091] Clause 5. The system of Clause 4, wherein the instructions further cause the electronic device to determine, based on the plurality of targets and the first and second subsets of the plurality of predictions, (i) a first performance metric corresponding to a performance of the ML model with respect to the first subset of the plurality of predictions, and (ii) a secondperformance metric corresponding to the performance of the ML model with respect to the second subset of the plurality’ of predictions.

[0092] Clause 6. The system of Clause 5, wherein the instructions further cause the electronic device to: determine (i) that the fairness metric does not satisfy a first fairness threshold or (ii) that a difference between the first and second performance metrics does not satisfy a second fairness threshold; and trigger an alert responsive to determining (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold, the alert indicating (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold.

[0093] Clause 7. The system of either Clause 5 or Clause 6, wherein the instructions further cause the electronic device to: generate a graphic representation of the first and second performance metrics, the graphic representation comprising a bar graph, a gauge chart, or a table; and present the graphic representation via the display device; wherein the first and second performance metrics each comprise: (i) a respective accuracy metric, (ii) a respective AUC metric, (iii) a respective log loss metric, (iv) a respective precision metric, (v) a respective recall metric, and (vi) a respective Fl score; or (i) a respective mean squared error metric, (ii) a respective mean absolute error metric, (iii) a respective root mean square error metric, (iv) a respective median absolute deviation metric, (v) a respective mean absolute percentage error metric, (vi) a respective mean prediction metric, (vii) a respective mean overprediction metric, (viii) a respective mean underprediction metric, (ix) a respective coefficient of determination, (x) a respective Pearson correlation metric, and (xi) a respective Spearman correlation metric.

[0094] Clause 8. The system of any one of Clauses 4 through 7, wherein: the instructions further cause the electronic device to generate a graphic representation of the fairness metric, the graphic representation comprising a bar graph, a gauge chart, or a table; presenting the fairness metric comprises presenting the graphic representation; and the fairness metric comprises: (i) an average odds difference metric, (ii) a disparate impact metric, (iii) a statistical parity difference metric, (iv) an equal opportunity difference metric, (v) a Theil index metric, (vi) a true negative rate ratio, (vii) a false negative rate ratio, (viii) a false discovery rate ratio, (ix) a false omission rate ratio, (x) a true positive rate ratio, (xi) a false positive rate ratio, (xii) a precision ratio, (xiii) a predicted positive rate ratio, (xiv) a predicted prevalence rate ratio, and (xv) a negative predicted value ratio; or (i) a demographic parity difference metric, (ii) a directdensity measure of independence, (iii) a direct density measure of separation, and (iv) a direct density measure of sufficiency.

[0095] Clause 9. The system of any one of Clauses 2 through 8, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective credit score, a respective annual income, and a respective debt-to-income ratio; the particular criterion comprises eligibility for a particular loan amount; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality’ of example entity, a respective example credit score, a respective example annual income, and a respective example debt-to- income ratio and (ii) that indicates whether each example entity is eligible for the particular loan amount.

[0096] Clause 10. The system of any one of Clauses 2 through 8, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective education indicator and a respective amount of work experience; the particular criterion comprises eligibility for an interview: and the ML model is trained using training data (i) that comprises, for each example entity’ of a plurality’ of example entities, a respective example education indicator and a respective example amount of work experience and (ii) that indicates whether each example entity is eligible for an interview.

[0097] Clause 11. The system of any one of Clauses 2 through 10. wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; and the particular protected characteristic comprises a particular type of protected characteristic comprising age, disability status, gender, marital status, maternity status, national origin, race, religion, sex, sexual orientation, or socioeconomic status.

[0098] Clause 12. A computer-implemented method for determining fairness of an ML model with respect to entities that have a protected characteristic, the method comprising: procuring an ML model configured to (i) receive data regarding each entity of a plurality of entities and (ii) provide, for each entity, a respective prediction regarding whether the respective entity’ meets a particular criterion and a respective confidence score corresponding to the respective prediction; receiving, from the ML model, (i) a plurality of predictions comprising the respective prediction for each entity of a plurality of entities and (ii) a plurality of confidence scores comprising the respective confidence score for each entity; receiving,from a user input, a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic; designating, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions comprising the respective prediction for each entity' that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic; determining a fairness metric for the ML model based on the first and second subsets of the plurality' of predictions, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; and presenting the fairness metric via a display device.

[0099] Clause 13. The computer-implemented method of Clause 12, wherein the ML model is trained using training data (i) that regards each example entity of a plurality of example entities and (ii) that indicates whether each example entity meets the particular criterion.

[0100] Clause 14. The computer-implemented method of either Clause 12 or Clause 13, further comprising receiving a plurality7of targets from the user input, the plurality7of targets comprising a respective target for each entity that indicates whether the respective entity meets the particular criterion, and wherein determining the fairness metric is further based on the plurality of targets.

[0101] Clause 15. The computer-implemented method of Clause 14, further comprising determining, based on the plurality of targets and the first and second subsets of the plurality of predictions, (i) a first performance metric corresponding to a performance of the ML model with respect to the first subset of the plurality of predictions, and (ii) a second performance metric corresponding to the performance of the ML model with respect to the second subset of the plurality of predictions.

[0102] Clause 16. The computer-implemented method of Clause 15, further comprising: determining (i) that the fairness metric does not satisfy a first fairness threshold or (ii) that a difference between the first and second performance metrics does not satisfy7a second fairness threshold; and triggering an alert responsive to determining (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy the second fairnessthreshold, the alert indicating (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy’ the second fairness threshold.

[0103] Clause 17. The computer-implemented method of either Clause 15 or Clause 16, further comprising: generating a graphic representation of the first and second performance metrics, the graphic representation comprising a bar graph, a gauge chart, or a table; and presenting the graphic representation via the display device; wherein the first and second performance metrics each comprise: (i) a respective accuracy metric, (ii) a respective AUC metric, (iii) a respective log loss metric, (iv) a respective precision metric, (v) a respective recall metric, and (vi) a respective Fl score; or (i) a respective mean squared error metric, (ii) a respective mean absolute error metric, (iii) a respective root mean square error metric, (iv) a respective median absolute deviation metric, (v) a respective mean absolute percentage error metric, (vi) a respective mean prediction metric, (vii) a respective mean overprediction metric, (viii) a respective mean underprediction metric, (ix) a respective coefficient of determination, (x) a respective Pearson correlation metric, and (xi) a respective Spearman correlation metric.

[0104] Clause 18. The computer-implemented method of any one of Clauses 14 through17, further comprising: generating a graphic representation of the fairness metric, the graphic representation comprising a bar graph, a gauge chart, or a table; wherein presenting the fairness metric comprises presenting the graphic representation; and the fairness metric comprises: (i) an average odds difference metric, (ii) a disparate impact metric, (iii) a statistical parity difference metric, (iv) an equal opportunity difference metric, (v) a Theil index metric, (vi) a true negative rate ratio, (vii) a false negative rate ratio, (viii) a false discovery rate ratio, (ix) a false omission rate ratio, (x) a true positive rate ratio, (xi) a false positive rate ratio, (xii) a precision ratio, (xiii) a predicted positive rate ratio, (xiv) a predicted prevalence rate ratio, and (xv) a negative predicted value ratio; or (i) a demographic parity difference metric, (ii) a direct densify measure of independence, (iii) a direct densify’ measure of separation, and (iv) a direct densify measure of sufficiency.

[0105] Clause 19. The computer-implemented method of any one of Clauses 12 through18, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective credit score, a respective annual income, and a respective debt-to-income ratio; the particular criterion comprises eligibility for a particular loan amount; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality’ of example entities, a respective example credit score, a respective example annual income, and arespective example debt-to-income ratio and (ii) that indicates whether each example entity is eligible for the particular loan amount.

[0106] Clause 20. The computer-implemented method of any one of Clauses 12 through 18, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective education indicator and a respective amount of work experience; the particular criterion comprises eligibility for an interview; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality of example entities, a respective example education indicator and a respective example amount of work experience and (ii) that indicates whether each example entity is eligible for an interview.

[0107] Further Consideration:

[0108] It is understood that the specific order or hierarchy of steps in the processes disclosed is an illustration of example approaches. Based upon design preferences, it is understood that the specific order or hierarchy of steps in the processes may be rearranged. Some of the steps may be performed simultaneously. The accompanying method claims present elements of the various steps in a sample order, and are not meant to be limited to the specific order or hierarchy presented.

[0109] The previous description is provided to enable any person skilled in the art to practice the various aspects described herein. The previous description provides various examples of the subject technology7, and the subject technology7is not limited to these examples. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects.

[0110] Thus, the claims are not intended to be limited to the aspects shown herein but are to be accorded the full scope consistent with the language of the claims. For example, reference to an element in the singular is not intended to mean “one and only one” unless specifically so stated, but rather “one or more.” Moreover, unless specifically stated otherwise, the term “some” refers to one or more. Pronouns in the masculine (e.g., his) include the feminine and neuter gender (e.g., her and its) and vice versa. Headings and subheadings, if any, are used for convenience only and do not limit the invention described herein.

Claims

WHAT IS CLAIMED IS:

1. A system for improving fairness of an ML model with respect to entities and protected characteristics, the system comprising: an ML model (i) trained using training data that regards each example entity of a plurality of example entities and that indicates whether each example entity meets a particular criterion and (ii) configured to receive data regarding each entity of a plurality of entities and provide, for each entity, a respective prediction regarding whether the respective entity meets the particular criterion and a respective confidence score corresponding to the respective prediction; and an electronic device comprising a display, a processor, and a non-transitory. computer- readable storage medium storing instructions that, when executed by the processor, cause the electronic device to: provide the data to the ML model; receive, from the ML model, (i) a plurality of predictions comprising the respective prediction for each entity and (ii) a plurality of confidence scores comprising the respective confidence score for each entity7; receive, from a user input, (i) a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic and (ii) a plurality of targets comprising a respective target for each entity that indicates whether the respective entity meets the particular criterion; designate, based on the plurality' of protected characteristic indicators, (i) a first subset of the plurality of predictions comprising the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic; determine a fairness metric for the ML model based on (i) the first and second subsets of the plurality of predictions, (ii) the plurality of confidence scores, and (fii) the plurality of targets, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; determine that the fairness metric does not satisfy a fairness threshold; anddisplay a notice via the display responsive to determining that the fairness metric does not satisfy the fairness threshold, the notice indicating that the fairness metric does not satisfy the fairness threshold.

2. An electronic device for improving fairness of an ML model with respect to entities and protected characteristics, the electronic device comprising a processor and anon-transitory, computer-readable storage medium storing instructions that, when executed by the processor, cause the electronic device to: receive, from an ML model, (i) a plurality of predictions comprising a respective prediction for each entity of a plurality of entities, the respective prediction regarding whether the respective entity meets a particular criterion and (ii) a plurality of confidence scores comprising a respective confidence score for each entity, the respective confidence score corresponding to the respective prediction, the ML model being configured to (i) receive data regarding each entity and (ii) provide, for each entity, the respective prediction and the respective confidence score; receive, from a user input, a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity that indicates whether the respective entity has a particular protected characteristic; designate, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions comprising the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic; determine a fairness metric for the ML model based on the first and second subsets of the plurality of predictions, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; and present the fairness metric via a display device associated with the electronic device.

3. The electronic device of Claim 2, wherein the ML model is trained using training data (i) that regards each example entity7of a plurality of example entities and (ii) that indicates whether each example entity meets the particular criterion.

4. The electronic device of Claim 2, wherein determining the fairness metric is further based on a plurality of targets comprising a respective target for each entity that indicates whether the respective entity meets the particular criterion, and the instructions further cause the electronic device to receive the plurality of targets from the user input.

5. The electronic device of Claim 4, wherein the instructions further cause the electronic device to determine, based on the plurality of targets and the first and second subsets of the plurality of predictions, (i) a first performance metric corresponding to a performance of the ML model with respect to the first subset of the plurality of predictions, and (ii) a second performance metric corresponding to the performance of the ML model with respect to the second subset of the plurality’ of predictions.

6. The electronic device of Claim 5, wherein the instructions further cause the electronic device to: determine (i) that the fairness metric does not satisfy a first fairness threshold or (ii) that a difference between the first and second performance metrics does not satisfy a second fairness threshold; and trigger an alert responsive to determining (i) that the fairness metric does not satisfy7the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold, the alert indicating (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold.

7. The electronic device of Claim 5, wherein the instructions further cause the electronic device to: generate a graphic representation of the first and second performance metrics, the graphic representation comprising a bar graph, a gauge chart, or a table; and present the graphic representation via the display device; wherein the first and second performance metrics each comprise:(i) a respective accuracy metric, (ii) a respective AUC metric, (iii) a respective log loss metric, (iv) a respective precision metric, (v) a respective recall metric, and (vi) a respective Fl score; or(i) a respective mean squared error metric, (ii) a respective mean absolute error metric, (iii) a respective root mean square error metric, (iv) a respective median absolute deviation metric, (v) a respective mean absolute percentage error metric, (vi) arespective mean prediction metric, (vii) a respective mean overprediction metric, (viii) a respective mean underprediction metric, (ix) a respective coefficient of determination, (x) a respective Pearson correlation metric, and (xi) a respective Spearman correlation metric.

8. The electronic device of Claim 4, wherein: the instructions further cause the electronic device to generate a graphic representation of the fairness metric, the graphic representation comprising a bar graph, a gauge chart, or a table; presenting the fairness metric comprises presenting the graphic representation; and the fairness metric comprises:(i) an average odds difference metric, (ii) a disparate impact metric, (iii) a statistical parity difference metric, (iv) an equal opportunity difference metric, (v) a Theil index metric, (vi) a true negative rate ratio, (vii) a false negative rate ratio, (viii) a false discovery rate ratio, (ix) a false omission rate ratio, (x) a true positive rate ratio, (xi) a false positive rate ratio, (xii) a precision ratio, (xiii) a predicted positive rate ratio, (xiv) a predicted prevalence rate ratio, and (xv) a negative predicted value ratio; or(i) a demographic parity difference metric, (ii) a direct density measure of independence, (iii) a direct density measure of separation, and (iv) a direct density measure of sufficiency.

9. The electronic device of Claim 2, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective credit score, a respective annual income, and a respective debt-to-income ratio; the particular criterion comprises eligibility7for a particular loan amount; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality of example entity, a respective example credit score, a respective example annual income, and a respective example debt-to-income ratio and (ii) that indicates whether each example entity is eligible for the particular loan amount.

10. The electronic device of Claim 2, wherein: the plurality of entities comprises a plurality of people;each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective education indicator and a respective amount of work experience; the particular criterion comprises eligibility for an interview; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality of example entities, a respective example education indicator and a respective example amount of work experience and (ii) that indicates whether each example entity is eligible for an interview.

11. The electronic device of Claim 2, wherein: the plurality of entities comprises a plurality of people; each respective entity' comprises a respective person; and the particular protected characteristic comprises a particular type of protected characteristic comprising age, disability status, gender, marital status, maternity status, national origin, race, religion, sex, sexual orientation, or socioeconomic status.

12. A computer-implemented method for improving fairness of an ML model with respect to entities and protected characteristics, the method comprising: procuring an ML model configured to (i) receive data regarding each entity of a plurality of entities and (ii) provide, for each entity, a respective prediction regarding whether the respective entity meets a particular criterion and a respective confidence score corresponding to the respective prediction; receiving, from the ML model, (i) a plurality' of predictions comprising the respective prediction for each entity of a plurality of entities and (ii) a plurality’ of confidence scores comprising the respective confidence score for each entity; receiving, from a user input, a plurality of protected characteristic indicators comprising a respective protected characteristic indicator for each entity’ that indicates whether the respective entity has a particular protected characteristic; designating, based on the plurality of protected characteristic indicators, (i) a first subset of the plurality of predictions comprising the respective prediction for each entity that has the particular protected characteristic and (ii) a second subset of the plurality of predictions comprising the respective prediction for each entity that does not have the particular protected characteristic;determining a fairness metric for the ML model based on the first and second subsets of the plurality of predictions, the fairness metric indicating whether the ML model provides fair predictions with respect to entities that have the particular protected characteristic and entities that do not have the particular protected characteristic; and presenting the fairness metric via a display device.

13. The computer-implemented method of Claim 12, wherein the ML model is trained using training data (i) that regards each example entity of a plurality of example entities and (ii) that indicates whether each example entity meets the particular criterion.

14. The computer-implemented method of Claim 12, further comprising receiving a plurality of targets from the user input, the plurality of targets comprising a respective target for each entity that indicates whether the respective entity meets the particular criterion, and wherein determining the fairness metric is further based on the plurality of targets.

15. The computer-implemented method of Claim 14, further comprising determining, based on the plurality of targets and the first and second subsets of the plurality of predictions,(i) a first performance metric corresponding to a performance of the ML model with respect to the first subset of the plurality of predictions, and (ii) a second performance metric corresponding to the performance of the ML model with respect to the second subset of the plurality of predictions.

16. The computer-implemented method of Claim 1 , further comprising: determining (i) that the fairness metric does not satisfy a first fairness threshold or(ii) that a difference between the first and second performance metrics does not satisfy a second fairness threshold; and triggering an alert responsive to determining (i) that the fairness metric does not satisfy7the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold, the alert indicating (i) that the fairness metric does not satisfy the first fairness threshold or (ii) that the difference does not satisfy the second fairness threshold.

17. The computer-implemented method of Claim 1 , further comprising: generating a graphic representation of the first and second performance metrics, the graphic representation comprising a bar graph, a gauge chart, or a table; andpresenting the graphic representation via the display device; wherein the first and second performance metrics each comprise:(i) a respective accuracy metric, (ii) a respective AUC metric, (iii) a respective log loss metric, (iv) a respective precision metric, (v) a respective recall metric, and (vi) a respective Fl score; or(i) a respective mean squared error metric, (ii) a respective mean absolute error metric, (iii) a respective root mean square error metric, (iv) a respective median absolute deviation metric, (v) a respective mean absolute percentage error metric, (vi) a respective mean prediction metric, (vii) a respective mean overprediction metric, (viii) a respective mean underprediction metric, (ix) a respective coefficient of determination, (x) a respective Pearson correlation metric, and (xi) a respective Spearman correlation metric.

18. The computer-implemented method of Claim 14, further comprising: generating a graphic representation of the fairness metric, the graphic representation comprising a bar graph, a gauge chart, or a table; wherein presenting the fairness metric comprises presenting the graphic representation; and wherein the fairness metric comprises:(i) an average odds difference metric, (ii) a disparate impact metric, (iii) a statistical parity difference metric, (iv) an equal opportunity difference metric, (v) a Theil index metric, (vi) a true negative rate ratio, (vii) a false negative rate ratio, (viii) a false discovery rate ratio, (ix) a false omission rate ratio, (x) a true positive rate ratio, (xi) a false positive rate ratio, (xii) a precision ratio, (xiii) a predicted positive rate ratio, (xiv) a predicted prevalence rate ratio, and (xv) a negative predicted value ratio; or(i) a demographic parity difference metric, (ii) a direct density' measure of independence, (iii) a direct density measure of separation, and (iv) a direct density measure of sufficiency.

19. The computer-implemented method of Claim 12, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective credit score, a respective annual income, and a respective debt-to-income ratio;the particular criterion comprises eligibility for a particular loan amount; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality of example entities, a respective example credit score, a respective example annual income, and a respective example debt-to-income ratio and (ii) that indicates whether each example entity is eligible for the particular loan amount.

20. The computer-implemented method of Claim 12, wherein: the plurality of entities comprises a plurality of people; each respective entity comprises a respective person; the data comprises, for each entity of the plurality of entities, a respective education indicator and a respective amount of work experience; the particular criterion comprises eligibility for an interview; and the ML model is trained using training data (i) that comprises, for each example entity of a plurality of example entities, a respective example education indicator and a respective example amount of work experience and (ii) that indicates whether each example entity is eligible for an interview.

Images