Fair AI Tools Action Set

Mitigating Bias of a Neural Network Model

This section contains PROC CAS code.

Note: Input data must be accessible in your CAS session, either as a CAS table or as a transient-scope table. A CAS table has a two-level name: the first level is your CAS engine libref, and the second level is the table name. You refer to this table in the CAS procedure by specifying only the second level. For more information about two-level names, see Chapter 2, Shared Concepts (SAS Viya: Machine Learning Procedures). A transient-scope table is called directly from the action and exists in memory for the duration of the action. For more information about accessing data, see SAS Viya: System Programming Guide. For more information about PROC CAS and programming in CASL, see SAS Cloud Analytic Services: CASL Programmer’s Guide and SAS Cloud Analytic Services: CASL Reference.

This example trains a neural network model in order to reduce bias according to the demographic parity metric by using the home equity data, which are available in the sampsio.hmeq data set. The data set has a binary target named BAD, and the example uses the variable JOB as the sensitive attribute.

You can load the sampsio.hmeq data set into your CAS session by specifying your CAS engine libref in the SET statement in the following DATA step:

data mycas.hmeq;
    set sampsio.hmeq;
    where JOB ne '';
run;

These statements assume that your CAS engine libref is named mycas, but you can substitute any appropriately defined CAS engine libref.

The following statements use the mitigateBiasNeuralNet action to run the exponentiated gradient reduction (EGR) algorithm in order to train a model that has reduced demographic parity bias, which is calculated on the sensitive attribute JOB. Note that the syntax of the trainOptions parameter here is the same as the syntax of the neuralNet.annTrain action.

proc cas;
fairAITools.mitigateBiasNeuralNet /
    mitigateOptions={biasMetric="DEMOGRAPHICPARITY",
                     predictedVariables={"P_BAD1", "P_BAD0"},
                     responseLevels={"1", "0"},
                     seed="12345",
                     sensitiveVariable="JOB"},
    trainOptions={combs={"LINEAR"},
                  encodeName="True",
                  errorFunc="ENTROPY",
                  hiddens={2},
                  inputs={"CLAGE",
                         "DEBTINC",
                         "DEROG",
                         "JOB",
                         "LOAN",
                         "MORTDUE",
                         "REASON",
                         "VALUE"},
                  nominals={"BAD",
                         "JOB",
                         "REASON"},
                  saveState={name="hmeq_nnet_astore",
                         replace="True"},
                  seed="54321",
                  std="MIDRANGE",
                  table="hmeq",
                  target="BAD",
                  targetAct="SOFTMAX"};
run;

The mitigateOptions parameter includes the following subparameters:

The biasMetric subparameter specifies that the demographic parity be used as the bias measurement.
The predictedVariables subparameter specifies that the variables P_BAD0 and P_BAD1 be used as the model predictions. A predicted variable is given for each level of the target variable.
The responseLevels subparameter specifies the levels of the target that correspond to each variable that is listed in the predictedVariables subparameter. The order of the response levels must match the order of the predicted variables.
The seed subparameter specifies the seed to use for pseudorandom number generation.
The sensitiveVariable subparameter specifies that the variable JOB be used as the sensitive variable in assessing and mitigating bias.

The trainOptions parameter includes the following subparameters that are passed to the neuralNet.annTrain action call:

The combs subparameter specifies the LINEAR combination function for the neurons in each hidden layer.
The encodeName subparameter specifies that the variable names of the predicted probabilities of the target be encoded.
The errorFunc subparameter specifies the error function to train the network. ENTROPY is the default setting for nominal targets.
The hiddens subparameter specifies the number of hidden neurons for each hidden layer in the neural network feedforward model.
The inputs subparameter specifies that the variables CLAGE, DEBTINC, DEROG, JOB, LOAN, MORTDUE, REASON, and VALUE be used as inputs.
The nominals subparameter specifies that the variables BAD, JOB, and REASON be used as nominal variables.
The saveState subparameter names the data table for saving the analytic store.
The seed subparameter specifies the seed to use to perform sampling and partitioning tasks.
The std subparameter specifies the standardization to use on the interval variables.
The table subparameter names the input data table for training.
The target subparameter specifies that the variable BAD be used as the target variable.
The targetAct subparameter specifies the SOFTMAX function as the activation function for the neurons in the output layer.

The "Mitigation Information" table that these statements produce is shown in Output 15.15.1. The table shows some key information about the model and values of the key hyperparameters that the EGR algorithm uses.

The "Iteration History" table that these statements produce is shown in Output 15.15.2. The table displays the iteration history of the overall demographic parity, the overall misclassification rate, and the overall multiclass log loss. From the table, you can see that without bias mitigation, the demographic parity gap between those two sensitive attributes is about 3.7%, and after seven iterations of the EGR algorithm it has been reduced to about 0.2%. Meanwhile, the overall accuracy of the model increases slightly.

The "Group Metrics" table that these statements produce is shown in Output 15.15.3. The table displays the model assessment statistics and average prediction for each sensitive attribute during each iteration of the EGR algorithm.

Output 15.15.1: Mitigation Information

Results from fairAITools.mitigateBiasNeuralNet

Mitigation Information
Description	Character Value	Value
Bias Metric	DEMOGRAPHICPARITY	.
Sensitive Variable	JOB	.
Response	BAD	.
Bound Value	100	100
Learning Rate	0.01	0.01
Maximum Iteration	10	10
Tolerance	0.005	0.005
Event Level	1	.

Output 15.15.2: Iteration History

EGR Iteration History
Iteration	Demographic Parity (Statistical Parity)	Misclassification Rate	Multi-Class Log Loss
1	0.036697	0.0863	0.2861
2	0.018349	0.0879	0.2897
3	0.018349	0.0874	0.2889
4	0.018349	0.0868	0.2884
5	0.018349	0.0868	0.2879
6	0.018349	0.0874	0.2888
7	0.002094	0.0877	0.2906

Output 15.15.3: Group Metrics History

Group Metrics
Iteration	JOB	Number of Observations	True Positive Rate	False Positive Rate	Area Under ROC	Misclassification Rate	Multi-Class Log Loss	Best Kolmogorov-Smirnov Along ROC	Cumulative Lift	Cumulative Response	Proportion Into Event Level	Average Prediction for Event
1	Mgr	471	0.0426	0.0000	0.6520	0.0955	0.2943	0.2862	3.4043	34.0426	0.0026	0.1137
	Office	618	0.0000	0.0000	0.5335	0.0631	0.2274	0.0666	4.3590	43.5897	0.0000	0.0669
	Other	1410	0.0548	0.0040	0.5476	0.1014	0.3299	0.0913	2.1918	21.9178	0.0054	0.0951
	ProfExe	925	0.0167	0.0000	0.5495	0.0638	0.2302	0.0988	2.6667	26.6667	0.0008	0.0675
	Sales	64	0.2500	0.0000	0.6204	0.1875	0.5266	0.3125	3.1250	31.2500	0.0367	0.1309
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3418	0.0714	4.2857	42.8571	0.0000	0.0677
2	Mgr	471	0.0000	0.0000	0.6352	0.0998	0.3075	0.2650	3.4043	34.0426	0.0000	0.0827
	Office	618	0.0000	0.0000	0.5256	0.0631	0.2304	0.0513	3.3333	33.3333	0.0000	0.0682
	Other	1410	0.0411	0.0040	0.5749	0.1028	0.3262	0.0982	1.7808	17.8082	0.0046	0.1119
	ProfExe	925	0.0167	0.0000	0.5411	0.0638	0.2349	0.0822	2.1667	21.6667	0.0008	0.0677
	Sales	64	0.1250	0.0000	0.6250	0.2188	0.6017	0.2500	2.5000	25.0000	0.0183	0.0984
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3501	0.0714	4.2857	42.8571	0.0000	0.0674
3	Mgr	471	0.0426	0.0000	0.6410	0.0955	0.3007	0.2697	3.4043	34.0426	0.0026	0.0914
	Office	618	0.0000	0.0000	0.5256	0.0631	0.2289	0.0513	3.5897	35.8974	0.0000	0.0665
	Other	1410	0.0411	0.0040	0.5485	0.1028	0.3291	0.0905	1.9863	19.8630	0.0046	0.1179
	ProfExe	925	0.0167	0.0000	0.5411	0.0638	0.2331	0.0822	2.5000	25.0000	0.0008	0.0662
	Sales	64	0.1250	0.0000	0.6250	0.2188	0.5848	0.2500	3.1250	31.2500	0.0183	0.1035
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3475	0.0714	4.2857	42.8571	0.0000	0.0659
4	Mgr	471	0.0426	0.0000	0.6910	0.0955	0.2987	0.3573	3.4043	34.0426	0.0026	0.0988
	Office	618	0.0000	0.0000	0.5256	0.0631	0.2284	0.0513	3.5897	35.8974	0.0000	0.0652
	Other	1410	0.0548	0.0040	0.5949	0.1014	0.3291	0.1610	2.1233	21.2329	0.0054	0.1124
	ProfExe	925	0.0167	0.0000	0.5411	0.0638	0.2324	0.0822	2.6667	26.6667	0.0008	0.0651
	Sales	64	0.1250	0.0000	0.6250	0.2188	0.5893	0.2500	3.1250	31.2500	0.0183	0.0999
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3464	0.0714	4.2857	42.8571	0.0000	0.0650
5	Mgr	471	0.0426	0.0000	0.6352	0.0955	0.2987	0.2862	3.4043	34.0426	0.0026	0.1137
	Office	618	0.0000	0.0000	0.5464	0.0631	0.2276	0.0922	4.8718	48.7179	0.0000	0.0655
	Other	1410	0.0548	0.0040	0.6059	0.1014	0.3291	0.2023	2.2603	22.6027	0.0054	0.1078
	ProfExe	925	0.0167	0.0000	0.5494	0.0638	0.2314	0.0988	2.6667	26.6667	0.0008	0.0655
	Sales	64	0.1250	0.0000	0.6563	0.2188	0.5898	0.3125	3.1250	31.2500	0.0183	0.0984
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3432	0.0714	4.2857	42.8571	0.0000	0.0658
6	Mgr	471	0.0426	0.0000	0.6874	0.0955	0.2996	0.3268	3.4043	34.0426	0.0026	0.1007
	Office	618	0.0000	0.0000	0.5256	0.0631	0.2290	0.0513	4.3590	43.5897	0.0000	0.0644
	Other	1410	0.0411	0.0040	0.5946	0.1028	0.3280	0.1821	2.2603	22.6027	0.0046	0.1051
	ProfExe	925	0.0167	0.0000	0.5411	0.0638	0.2328	0.0822	2.8333	28.3333	0.0008	0.0643
	Sales	64	0.1250	0.0000	0.6250	0.2188	0.6174	0.2500	3.1250	31.2500	0.0183	0.0876
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3466	0.0714	4.2857	42.8571	0.0000	0.0646
7	Mgr	471	0.0000	0.0000	0.6309	0.0998	0.3053	0.2484	3.4043	34.0426	0.0000	0.0866
	Office	618	0.0000	0.0000	0.5256	0.0631	0.2309	0.0513	4.1026	41.0256	0.0000	0.0647
	Other	1410	0.0342	0.0000	0.5488	0.1000	0.3260	0.0934	1.8493	18.4932	0.0021	0.0924
	ProfExe	925	0.0000	0.0000	0.5167	0.0649	0.2351	0.0333	2.8333	28.3333	0.0000	0.0642
	Sales	64	0.0000	0.0000	0.6250	0.2500	0.6590	0.2500	3.1250	31.2500	0.0000	0.0758
	Self	128	0.0000	0.0000	0.5357	0.1094	0.3511	0.0714	5.7143	57.1429	0.0000	0.0646

Mitigating Bias of a Neural Network Model

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 illustrates the mitigateBiasNeuralNet action by using the home equity data that are described in the CASL version of the example.

mitigateOptions={biasMetric="DEMOGRAPHICPARITY",
                 predictedVariables={"P_BAD1", "P_BAD0"},
                 responseLevels={"1", "0"},
                 seed=12345,
                 sensitiveVariable="JOB"},
trainOptions={combs={"LINEAR"},
              encodeName=True,
              errorFunc="ENTROPY",
              hiddens={2},
              inputs={"CLAGE",
                      "DEBTINC",
                      "DEROG",
                      "JOB",
                      "LOAN",
                      "MORTDUE",
                      "REASON",
                      "VALUE"},
              nominals={"BAD",
                        "JOB",
                        "REASON"},
              saveState={name="hmeq_nnet_astore",
                         replace=True},
              seed=54321,
              std="MIDRANGE",
              table="hmeq",
              target="BAD",
              targetAct="SOFTMAX"
}

The mitigateOptions parameter includes the following subparameters:

The biasMetric subparameter specifies that the demographic parity be used as the bias measurement.
The predictedVariables subparameter specifies that the variables P_BAD0 and P_BAD1 be used as the model predictions. A predicted variable is given for each level of the target variable.
The responseLevels subparameter specifies the levels of the target that correspond to each variable that is listed in the predictedVariables subparameter. The order of the response levels must match the order of the predicted variables.
The seed subparameter specifies the seed to use for pseudorandom number generation.
The sensitiveVariable subparameter specifies that the variable JOB be used as the sensitive variable in assessing and mitigating bias.

The trainOptions parameter includes the following subparameters that are passed to the neuralNet.annTrain action call:

The combs subparameter specifies the LINEAR combination function for the neurons in each hidden layer.
The encodeName subparameter specifies that the variable names of the predicted probabilities of the target be encoded.
The errorFunc subparameter specifies the error function to train the network. ENTROPY is the default setting for nominal targets.
The hiddens subparameter specifies the number of hidden neurons for each hidden layer in the neural network feedforward model.
The inputs subparameter specifies that the variables CLAGE, DEBTINC, DEROG, JOB, LOAN, MORTDUE, REASON, and VALUE be used as inputs.
The nominals subparameter specifies that the variables BAD, JOB, and REASON be used as nominal variables.
The saveState subparameter names the data table for saving the analytic store.
The seed subparameter specifies the seed to use to perform sampling and partitioning tasks.
The std subparameter specifies the standardization to use on the interval variables.
The table subparameter names the input data table for training.
The target subparameter specifies that the variable BAD be used as the target variable.
The targetAct subparameter specifies the SOFTMAX function as the activation function for the neurons in the output layer.

Mitigating Bias of a Neural Network Model

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 illustrates the mitigateBiasNeuralNet action by using the home equity data that are described in the CASL version of the example.

s.fairaitools.mitigatebiasneuralnet(
    mitigateOptions=dict(biasMetric='DEMOGRAPHICPARITY',
                         predictedVariables=['P_BAD1', 'P_BAD0'],
                         responseLevels=['1', '0'],
                         seed='12345',
                         sensitiveVariable='JOB'),
    trainOptions=dict(combs=['LINEAR'],
                      encodeName='True',
                      errorFunc='ENTROPY',
                      hiddens=[2],
                      inputs=['CLAGE',
                             'DEBTINC',
                             'DEROG',
                             'JOB',
                             'LOAN',
                             'MORTDUE',
                             'REASON',
                             'VALUE'],
                      nominals=['BAD',
                             'JOB',
                             'REASON'],
                      saveState=dict(name='hmeq_nnet_astore',
                             replace='True'),
                      seed='54321',
                      std='MIDRANGE',
                      table='hmeq',
                      target='BAD',
                      targetAct='SOFTMAX'))

The mitigateOptions parameter includes the following subparameters:

The biasMetric subparameter specifies that the demographic parity be used as the bias measurement.
The predictedVariables subparameter specifies that the variables P_BAD0 and P_BAD1 be used as the model predictions. A predicted variable is given for each level of the target variable.
The responseLevels subparameter specifies the levels of the target that correspond to each variable that is listed in the predictedVariables subparameter. The order of the response levels must match the order of the predicted variables.
The seed subparameter specifies the seed to use for pseudorandom number generation.
The sensitiveVariable subparameter specifies that the variable JOB be used as the sensitive variable in assessing and mitigating bias.

The trainOptions parameter includes the following subparameters that are passed to the neuralNet.annTrain action call:

The combs subparameter specifies the LINEAR combination function for the neurons in each hidden layer.
The encodeName subparameter specifies that the variable names of the predicted probabilities of the target be encoded.
The errorFunc subparameter specifies the error function to train the network. ENTROPY is the default setting for nominal targets.
The hiddens subparameter specifies the number of hidden neurons for each hidden layer in the neural network feedforward model.
The inputs subparameter specifies that the variables CLAGE, DEBTINC, DEROG, JOB, LOAN, MORTDUE, REASON, and VALUE be used as inputs.
The nominals subparameter specifies that the variables BAD, JOB, and REASON be used as nominal variables.
The saveState subparameter names the data table for saving the analytic store.
The seed subparameter specifies the seed to use to perform sampling and partitioning tasks.
The std subparameter specifies the standardization to use on the interval variables.
The table subparameter names the input data table for training.
The target subparameter specifies that the variable BAD be used as the target variable.
The targetAct subparameter specifies the SOFTMAX function as the activation function for the neurons in the output layer.

Mitigating Bias of a Neural Network Model

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 illustrates the mitigateBiasNeuralNet action by using the home equity data that are described in the CASL version of the example.

cas.fairAITools.mitigateBiasNeuralNet(s,
    mitigateOptions=list(biasMetric='DEMOGRAPHICPARITY',
                         predictedVariables=c('P_BAD1', 'P_BAD0'),
                         responseLevels=c('1', '0'),
                         seed='12345',
                         sensitiveVariable='JOB'),
    trainOptions=list(combs=c('LINEAR'),
                      encodeName='True',
                      errorFunc='ENTROPY',
                      hiddens=c(2),
                      inputs=c('CLAGE',
                              'DEBTINC',
                              'DEROG',
                              'JOB',
                              'LOAN',
                              'MORTDUE',
                              'REASON',
                              'VALUE'),
                      nominals=c('BAD',
                              'JOB',
                              'REASON'),
                      saveState=list(name='hmeq_nnet_astore',
                              replace='True'),
                      seed='54321',
                      std='MIDRANGE',
                      table='hmeq',
                      target='BAD',
                      targetAct='SOFTMAX'))

The mitigateOptions parameter includes the following subparameters:

The biasMetric subparameter specifies that the demographic parity be used as the bias measurement.
The predictedVariables subparameter specifies that the variables P_BAD0 and P_BAD1 be used as the model predictions. A predicted variable is given for each level of the target variable.
The responseLevels subparameter specifies the levels of the target that correspond to each variable that is listed in the predictedVariables subparameter. The order of the response levels must match the order of the predicted variables.
The seed subparameter specifies the seed to use for pseudorandom number generation.
The sensitiveVariable subparameter specifies that the variable JOB be used as the sensitive variable in assessing and mitigating bias.

The trainOptions parameter includes the following subparameters that are passed to the neuralNet.annTrain action call:

The combs subparameter specifies the LINEAR combination function for the neurons in each hidden layer.
The encodeName subparameter specifies that the variable names of the predicted probabilities of the target be encoded.
The errorFunc subparameter specifies the error function to train the network. ENTROPY is the default setting for nominal targets.
The hiddens subparameter specifies the number of hidden neurons for each hidden layer in the neural network feedforward model.
The inputs subparameter specifies that the variables CLAGE, DEBTINC, DEROG, JOB, LOAN, MORTDUE, REASON, and VALUE be used as inputs.
The nominals subparameter specifies that the variables BAD, JOB, and REASON be used as nominal variables.
The saveState subparameter names the data table for saving the analytic store.
The seed subparameter specifies the seed to use to perform sampling and partitioning tasks.
The std subparameter specifies the standardization to use on the interval variables.
The table subparameter names the input data table for training.
The target subparameter specifies that the variable BAD be used as the target variable.
The targetAct subparameter specifies the SOFTMAX function as the activation function for the neurons in the output layer.

Last updated: August 04, 2026