Configuring fairness evaluations for machine learning models

If you log payload data when you prepare for model evaluations, you can configure fairness evaluations.

You can configure fairness evaluations manually or you can run a custom notebook to generate a configuration file. You can upload the configuration file to specify the settings for your evaluation.

When you configure fairness evaluations manually, you can specify the reference group (value) that you expect to represent favorable outcomes. You can also select the corresponding model attributes (features) to monitor for bias (for example, Age or Sex), that will be compared against the reference group. Depending on your training data, you can also specify the minimum and maximum sample size for evaluations.

Configuring fairness evaluations in watsonx.governance

When you evaluate prompt templates, you can review a summary of fairness evaluation results for the text classification tasks.

Supported fairness metrics

Supported languages: English only

When you enable fairness evaluations for machine learning models or generative AI assets, you can view a summary of evaluation results with metrics for the type of model that you're evaluating.

You can view the results of your fairness evaluations for machine learning modles on the Insights dashboard. For more information, see Reviewing fairness results.

The following metrics are supported by fairness evaluations:

Disparate impact

Disparate impact is specified as the fairness scores for different groups. Disparate impact compares the percentage of favorable outcomes for a monitored group to the percentage of favorable outcomes for a reference group.

• How it works: When you view the details of a model deployment, the Fairness section of the model summary that is displayed, provides the fairness scores for different groups that are described as metrics. The fairness scores are calculated with the disparate impact formula.

• Uses the confusion matrix to measure performance: No

• Do the math:

                    (num_positives(privileged=False) / num_instances(privileged=False))
Disparate impact =   ______________________________________________________________________
                    (num_positives(privileged=True) / num_instances(privileged=True))              

The num_positives value represents the number of individuals in the group who received a positive outcome, and the num_instances value represents the total number of individuals in the group. The privileged=False label specifies unprivileged groups and the privileged=True label specifies privileged groups. The positive outcomes are designated as the favorable outcomes, and the negative outcomes are designated as the unfavorable outcomes. The privileged group is designated as the reference group, and the unprivileged group is designated as the monitored group.

The calculation produces a percentage that specifies how often the rate that the unprivileged group receives the positive outcome is the same rate that the privileged group receives the positive outcome. For example, if a credit risk model assigns the “no risk” prediction to 80% of unprivileged applicants and to 100% of privileged applicants, that model has a disparate impact of 80%.

• Supported fairness details

  • The following details for fairness metrics are supported:
    • The favorable percentages for each of the groups
    • Fairness averages for all the fairness groups
    • Distribution of the data for each of the monitored groups
    • Distribution of payload data

Statistical parity difference

The statistical parity difference compares the percentage of favorable outcomes for monitored groups to reference groups.

• Description: Fairness metric that describes the fairness for the model predictions. It is the difference between the ratio of favorable outcomes in monitored and reference groups

  • Under 0: Higher benefits for the monitored group.
  • At 0: Both groups have equal benefit.
  • Over 0 Implies higher benefit for the reference group.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

                                    num_positives(privileged=False)     num_positives(privileged=True) 
Statistical parity difference =  ________________________________ -  ________________________________
                                    num_instances(privileged=False)     num_instances(privileged=True)

Impact score

The impact score compares the rate that monitored groups are selected to receive favorable outcomes to the rate that reference groups are selected to receive favorable outcomes.

• Do the math:

The following formula calculates the selection rate for each group:

                                      number of individuals receiving favorable outcomes
            Selection rate  =   ________________________________________________________
                                       total number of individuals

The following formula calculates the impact score:

                                      selection rate for monitored groups
           Impact score  =   ________________________________________________________
                                      selection rate for reference groups

• Thresholds:

  • Lower bound: 0.8
  • Upper bound: 1.0

• How it works: Higher scores indicate higher selection rates for monitored groups

False negative rate difference

The false negative rate difference gives the percentage of positive transactions that were incorrectly scored as negative by your model.

• Description: Returns the difference in false negative rates for the monitored and reference groups

  • At 0: Both groups have equal benefit.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating false negative rate (FNR):

                                        false negatives         
            False negative rate  =  __________________________
                                        all positives

The following formula is used for calculating false negative rate difference:

            False negative rate difference  =  FNR of monitored group - FNR of reference group

False positive rate difference

The false positive rate difference gives the percentage of negative transactions that were incorrectly scored as positive by your model.

• Description: Returns the ratio of false positive rate for the monitored group and reference groups.

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating false positive rate (FPR):

                                        false positives       
            False positive rate   =   ________________________
                                        total negatives

The following formula is used for calculating false positive rate difference:

            False positive rate difference  =   FPR of monitored group - FPR of reference group

False discovery rate difference

The false discovery rate difference gives the amount of false positive transactions as a percentage of all transactions with a positive outcome. It describes the pervasiveness of false positives among all positive transactions.

• Description: Returns the difference in false discovery rate for the monitored and reference groups.

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating the false discovery rate (FDR):

                                                false positives        
            False discovery rate  =   _________________________________________
                                        true positives + false positives

The following formula is used for calculating the false discovery rate difference:

            False discovery rate difference  = FDR of monitored group - FDR of reference group

False omission rate difference

The false omission rate difference gives the number of false negative transactions as a percentage of all transactions with a negative outcome. It describes the pervasiveness of false negatives among all negative transactions.

• Description: Returns the difference in false omission rate for the monitored and reference groups

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating the false omission rate (FOR):

                                                false negatives        
            False omission rate   =   ________________________________________
                                        true negatives + false negatives

The following formula is used for the false omission rate difference:

            False omission rate difference  =   FOR of monitored group - FOR of reference group                                         

Error rate difference

The error rate difference calculates the percentage of transactions that are incorrectly scored by your model.

• Description: Returns the difference in error rate for the monitored and reference groups.

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating the error rate (ER):

                                false positives + false negatives        
            Error rate  =   ___________________________________________
                                all positives + all negatives

The following formula is used for calculating the error rate difference:

            Error rate difference  = ER of monitored group - ER of reference group

Average odds difference

The average odds difference gives the percentage of transactions that was incorrectly scored by your model.

• Description: Returns the difference in error rate for the monitored and reference groups.

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating false positive rate (FPR):

                                        false positives       
            False positive rate   =   _________________________
                                        total negatives

The following formula is used for calculating true positive rate (TPR):

                                        True positives      
            True positive rate   =   ______________________
                                        All positives

The following formula is used for calculating average odds difference:

                                        (FPR monitored group - FPR reference group) + (TPR monitored group - TPR reference group)       
            Average odds difference  =   ___________________________________________________________________________________________

                                                                                    2

Average absolute odds difference

The average absolute odds difference compares the average of absolute difference in false positive rates and true positive rates between monitored groups and reference groups.

• Description: Returns the average of the absolute difference in false positive rate and true positive rate for the monitored and reference groups.

  • At 0: Both groups have equal odds.

• Uses the confusion matrix to measure performance: Yes

• Do the math:

The following formula is used for calculating false positive rate (FPR):

                                            false positives       
            False positive rate   =   ____________________________
                                            all negatives

The following formula is used for calculating true positive rate (TPR):

                                        True positives      
            True positive rate   =   ________________________
                                        All positives

The following formula is used for calculating average absolute odds difference:

                                                |FPR monitored group - FPR reference group| + |TPR monitored group - TPR reference group|      
            Average absolute odds difference  =   ______________________________________________________________________________________________

                                                                                            2