Fair AI Tools Action Set

This section contains Lua code for the analysis in the CASL version of this example, which contains details about the results.

For more information about coding in Lua, see Getting Started with SAS Viya for Lua and SAS Viya: System Programming Guide.

This example trains a model in order to reduce bias according to the demographic parity metric by using the data set simdata that is described in Example 15.1.

The following statements use the mitigateBias action to run the EGR algorithm in order to train a model with reduced demographic parity bias, calculated on the sensitive attribute mySensitiveVariable:

fairAITools.mitigateBias{
   table='simdata',
   predictedVariables={
      'P_nominalTargetA',
      'P_nominalTargetB',
      'P_nominalTargetC'
   },
   response='nominalTarget',
   responseLevels={'A', 'B', 'C'},
   event='A',
   sensitiveVariable='mySensitiveVariable',
   biasMetric='DEMOGRAPHICPARITY',
   learningRate=0.01,
   bound=200,
   tolerance=0.005,
   maxIters=10,
   cutoff=0.33,
   seed=12345,
   trainProgram= '
      decisionTree.gbtreeTrain result=train_res /
         table=table,
         weight=weight,
         target="nominalTarget",
         inputs={
            "linearPredictor",
            "absolutePredictor",
            "quadraticPredictor",
            "sinusoidPredictor"
         },
         nominals={"nominalTarget"},
         nBins=50,
         quantileBin=True,
         maxLevel=5,
         maxBranch=2,
         leafSize=5,
         missing="USEINSEARCH",
         minUseInSearch=1,
         binOrder=True,
         varImp=True,
         mergeBin=True,
         encodeName=True,
         nTree=500,
         seed=12345,
         ridge=1 ,
         savestate={
            name ="gb_astore",
            replace=True
         }
      ;
      astore.score result=score_res /
         table=table,
         rstore="gb_astore",
         casout=casout,
         copyVars=copyVars
       ;
   '
}

The table parameter names the input reference data table. The response parameter specifies the name of the target variable nominalTarget for the training model. The sensitiveVariable parameter specifies that the variable mySensitiveVariable be used as the sensitive variable in assessing bias and mitigating bias. The trainProgram parameter specifies the CASL program that contains the training and scoring code. The predictedVariables parameter specifies that the variables P_nominalTargetA, P_nominalTargetB, and P_nominalTargetC be used as the model predictions. A predicted variable is given for each level of the response variable. The responseLevels parameter specifies the levels of the target that correspond to each variable listed in the predictedVariables parameter. The order of the response levels must match the order of the predicted variables. The event parameter specifies that the level 'A' be used as the event level for the response variable. The cutoff parameter specifies the threshold of the prediction to be counted as the event. In this example, any observation whose prediction of P_nominalTargetA is greater than or equal to 0.33 is counted as an event. The biasMetric parameter specifies that the fairness measurement is demographic parity. The bound parameter specifies the value of the hyperparameter bound in the EGR algorithm. The learningRate parameter specifies the value of the learning rate in the EGR algorithm. If the tuneBound parameter is set to True, the learningRate parameter should be set to 0.01, because the bound value tuner uses the fixed learning rate of 0.01 initially. The maxIters parameter specifies the maximum number of iterations of the EGR algorithm. The seed parameter specifies the seed to use for pseudorandom number generation. The tolerance parameter specifies the threshold for constraint violations. Because constraint violations are calculated by using the difference in bias metrics between the individual sensitive attribute and the overall population, the tolerance value refers to the maximum difference in bias metrics between the individual sensitive attribute and the overall population. When the tolerance condition is satisfied, the EGR algorithm stops, and in this case, the number of iterations of the EGR algorithm might be less than the specified maxIters parameter value.

This section contains Python code for the analysis in the CASL version of this example, which contains details about the results.

For more information about coding in Python, see Getting Started with SAS Viya for Python and SAS Viya: System Programming Guide.

This example trains a model in order to reduce bias according to the demographic parity metric by using the data set simdata that is described in Example 15.1.

The following statements use the mitigateBias action to run the EGR algorithm in order to train a model with reduced demographic parity bias, calculated on the sensitive attribute mySensitiveVariable:

s.fairaitools.mitigatebias(
    biasMetric='DEMOGRAPHICPARITY',
    bound='200',
    cutoff='0.33',
    event='A',
    learningRate='0.01',
    maxIters='10',
    predictedVariables=['P_nominalTargetA', 'P_nominalTargetB',
                        'P_nominalTargetC'],
    response='nominalTarget',
    responseLevels=['A', 'B', 'C'],
    seed='12345',
    sensitiveVariable='mySensitiveVariable',
    table='simdata',
    tolerance='0.005',
    trainProgram='''
         decisionTree.gbtreeTrain result=train_res /
            table=table,
            weight=weight,
            target="nominalTarget",
            inputs={
               "linearPredictor",
               "absolutePredictor",
               "quadraticPredictor",
               "sinusoidPredictor"
            },
            nominals={"nominalTarget"},
            nBins=50,
            quantileBin=True,
            maxLevel=5,
            maxBranch=2,
            leafSize=5,
            missing="USEINSEARCH",
            minUseInSearch=1,
            binOrder=True,
            varImp=True,
            mergeBin=True,
            encodeName=True,
            nTree=500,
            seed=12345,
            ridge=1 ,
            savestate={
               name ="gb_astore",
               replace=True
            }
         ;
         astore.score result=score_res /
            table=table,
            rstore="gb_astore",
            casout=casout,
            copyVars=copyVars
          ;
      ''')

The table parameter names the input reference data table. The response parameter specifies the name of the target variable nominalTarget for the training model. The sensitiveVariable parameter specifies that the variable mySensitiveVariable be used as the sensitive variable in assessing bias and mitigating bias. The trainProgram parameter specifies the CASL program that contains the training and scoring code. The predictedVariables parameter specifies that the variables P_nominalTargetA, P_nominalTargetB, and P_nominalTargetC be used as the model predictions. A predicted variable is given for each level of the response variable. The responseLevels parameter specifies the levels of the target that correspond to each variable listed in the predictedVariables parameter. The order of the response levels must match the order of the predicted variables. The event parameter specifies that the level 'A' be used as the event level for the response variable. The cutoff parameter specifies the threshold of the prediction to be counted as the event. In this example, any observation whose prediction of P_nominalTargetA is greater than or equal to 0.33 is counted as an event. The biasMetric parameter specifies that the fairness measurement is demographic parity. The bound parameter specifies the value of the hyperparameter bound in the EGR algorithm. The learningRate parameter specifies the value of the learning rate in the EGR algorithm. If the tuneBound parameter is set to True, the learningRate parameter should be set to 0.01, because the bound value tuner uses the fixed learning rate of 0.01 initially. The maxIters parameter specifies the maximum number of iterations of the EGR algorithm. The seed parameter specifies the seed to use for pseudorandom number generation. The tolerance parameter specifies the threshold for constraint violations. Because constraint violations are calculated by using the difference in bias metrics between the individual sensitive attribute and the overall population, the tolerance value refers to the maximum difference in bias metrics between the individual sensitive attribute and the overall population. When the tolerance condition is satisfied, the EGR algorithm stops, and in this case, the number of iterations of the EGR algorithm might be less than the specified maxIters parameter value.

This section contains R code for the analysis in the CASL version of this example, which contains details about the results.

For more information about coding in R, see Getting Started with SAS Viya for R and SAS Viya: System Programming Guide.

This example trains a model in order to reduce bias according to the demographic parity metric by using the data set simdata that is described in Example 15.1.

The following statements use the mitigateBias action to run the EGR algorithm in order to train a model with reduced demographic parity bias, calculated on the sensitive attribute mySensitiveVariable:

cas.fairAITools.mitigateBias(s,
    biasMetric='DEMOGRAPHICPARITY',
    bound='200',
    cutoff='0.33',
    event='A',
    learningRate='0.01',
    maxIters='10',
    predictedVariables=c('P_nominalTargetA', 'P_nominalTargetB',
                         'P_nominalTargetC'),
    response='nominalTarget',
    responseLevels=c('A', 'B', 'C'),
    seed='12345',
    sensitiveVariable='mySensitiveVariable',
    table='simdata',
    tolerance='0.005',
    trainProgram='
         decisionTree.gbtreeTrain result=train_res /
            table=table,
            weight=weight,
            target="nominalTarget",
            inputs={
               "linearPredictor",
               "absolutePredictor",
               "quadraticPredictor",
               "sinusoidPredictor"
            },
            nominals={"nominalTarget"},
            nBins=50,
            quantileBin=True,
            maxLevel=5,
            maxBranch=2,
            leafSize=5,
            missing="USEINSEARCH",
            minUseInSearch=1,
            binOrder=True,
            varImp=True,
            mergeBin=True,
            encodeName=True,
            nTree=500,
            seed=12345,
            ridge=1 ,
            savestate={
               name ="gb_astore",
               replace=True
            }
         ;
         astore.score result=score_res /
            table=table,
            rstore="gb_astore",
            casout=casout,
            copyVars=copyVars
          ;
      ')

The table parameter names the input reference data table. The response parameter specifies the name of the target variable nominalTarget for the training model. The sensitiveVariable parameter specifies that the variable mySensitiveVariable be used as the sensitive variable in assessing bias and mitigating bias. The trainProgram parameter specifies the CASL program that contains the training and scoring code. The predictedVariables parameter specifies that the variables P_nominalTargetA, P_nominalTargetB, and P_nominalTargetC be used as the model predictions. A predicted variable is given for each level of the response variable. The responseLevels parameter specifies the levels of the target that correspond to each variable listed in the predictedVariables parameter. The order of the response levels must match the order of the predicted variables. The event parameter specifies that the level 'A' be used as the event level for the response variable. The cutoff parameter specifies the threshold of the prediction to be counted as the event. In this example, any observation whose prediction of P_nominalTargetA is greater than or equal to 0.33 is counted as an event. The biasMetric parameter specifies that the fairness measurement is demographic parity. The bound parameter specifies the value of the hyperparameter bound in the EGR algorithm. The learningRate parameter specifies the value of the learning rate in the EGR algorithm. If the tuneBound parameter is set to True, the learningRate parameter should be set to 0.01, because the bound value tuner uses the fixed learning rate of 0.01 initially. The maxIters parameter specifies the maximum number of iterations of the EGR algorithm. The seed parameter specifies the seed to use for pseudorandom number generation. The tolerance parameter specifies the threshold for constraint violations. Because constraint violations are calculated by using the difference in bias metrics between the individual sensitive attribute and the overall population, the tolerance value refers to the maximum difference in bias metrics between the individual sensitive attribute and the overall population. When the tolerance condition is satisfied, the EGR algorithm stops, and in this case, the number of iterations of the EGR algorithm might be less than the specified maxIters parameter value.