Business Context
A financial institution maintains multiple versions of a credit risk model. To avoid confusion in their model inventory, each generated scoring code needs a specific, human-readable model identifier. They are testing the `modelId` parameter and ensuring the process is robust to missing data in the training set.
About the Set : neuralNet
Training of classical artificial neural networks.
Discover all actions of neuralNet Data Preparation
Create a credit application dataset with applicant information. Intentionally introduce missing values (`.`) in the `Years_At_Address` variable to test model robustness.
Copied!
| 1 | DATA casuser.credit_applications; |
| 2 | call streaminit(789); |
| 3 | DO i = 1 to 1000; |
| 4 | Income = 30000 + rand('UNIFORM') * 150000; |
| 5 | Loan_Amount = 5000 + rand('UNIFORM') * 50000; |
| 6 | IF rand('UNIFORM') > 0.85 THEN Years_At_Address = . ; |
| 7 | ELSE Years_At_Address = rand('INTEGER', 0, 25); |
| 8 | Default_Flag = rand('BERNOULLI', 0.2); |
| 9 | OUTPUT; |
| 10 | END; |
| 11 | RUN; |
Étapes de réalisation
1
Train a neural network on the `credit_applications` data. The `annTrain` action will impute missing values by default. Store the model in `credit_risk_model_v2`.
Copied!
| 1 | PROC CAS; |
| 2 | neuralNet.annTrain / |
| 3 | TABLE={name='credit_applications', vars={{name='Years_At_Address', impute='MEAN'}}}, |
| 4 | inputs={{name='Income'}, {name='Loan_Amount'}, {name='Years_At_Address'}}, |
| 5 | target='Default_Flag', |
| 6 | casOut={name='credit_risk_model_v2', replace=true}; |
| 7 | RUN; |
2
Use `annCode` to generate scoring code, setting the `modelId` parameter to a specific version string: 'CreditRisk_NN_v2_Q42025'.
Copied!
| 1 | PROC CAS; |
| 2 | neuralNet.annCode / |
| 3 | modelTable={name='credit_risk_model_v2'}, |
| 4 | modelId='CreditRisk_NN_v2_Q42025', |
| 5 | code={casOut={name='credit_risk_scoring_code_v2', replace=true}}; |
| 6 | RUN; |
Expected Result
The `annCode` action should successfully generate scoring code. An inspection of the generated DATA step code must show that the output prediction variable is named using the custom ID (e.g., `P_CreditRisk_NN_v2_Q420251`). This confirms that the `modelId` parameter correctly customizes the output code for model management and versioning purposes, even when the training data contained missing values.