Business Context
A credit card company needs to process a very large dataset of transactions to build a fraud detection model. Due to the data volume and urgency, the model training must be efficient. This scenario tests the action's performance under load and its ability to leverage a distributed computing environment.
Data Preparation
Create a large dataset (1.5 million transactions) with features like transaction amount, time of day, and merchant category. The target 'Is_Fraud' is a rare event.
Copied!
| 1 | DATA transactions_large; |
| 2 | call streaminit(555); |
| 3 | DO i = 1 to 1500000; |
| 4 | Transaction_Amount = 5 + rand('Exponential', 50); |
| 5 | Time_Of_Day = floor(rand('Uniform') * 24); /* 0-23 hours */ |
| 6 | IF rand('Uniform') < 0.3 THEN Merchant_Category = 'Online'; |
| 7 | ELSE IF rand('Uniform') < 0.6 THEN Merchant_Category = 'Retail'; |
| 8 | ELSE Merchant_Category = 'Services'; |
| 9 | /* Fraud is more likely with high amounts and late at night */ |
| 10 | logit_p = -8 + (Transaction_Amount / 100) + (ifc(Time_Of_Day < 6, 1.5, 0)) + (ifc(Merchant_Category='Online', 0.5, 0)); |
| 11 | p = 1 / (1 + exp(-logit_p)); |
| 12 | Is_Fraud = rand('Bernoulli', p); |
| 13 | OUTPUT; |
| 14 | END; |
| 15 | RUN; |
| 16 | |
| 17 | DATA casuser.transactions_large; |
| 18 | SET transactions_large; |
| 19 | RUN; |
Étapes de réalisation
1
Load the large transaction dataset into CAS.
Copied!
| 1 | PROC CAS; |
| 2 | TABLE.promote TABLE={caslib='CASUSER', name='transactions_large'}; |
| 3 | RUN; |
| 4 | QUIT; |
2
Run bart.bartProbit using performance-oriented parameters: distribute chains across 4 workers, set a max training time, and enable in-memory options.
Copied!
| 1 | PROC CAS; |
| 2 | bart.bartProbit TABLE={name='transactions_large'}, |
| 3 | target='Is_Fraud', |
| 4 | inputs={'Transaction_Amount', 'Time_Of_Day'}, |
| 5 | nominals={'Merchant_Category'}, |
| 6 | nTree=250, /* More trees for complex interactions */ |
| 7 | nBI=500, |
| 8 | nMC=2500, |
| 9 | distributeChains=4, /* Simulate running on 4 workers */ |
| 10 | trainInMem=TRUE, /* Use more memory for speed */ |
| 11 | obsLeafMapInMem=TRUE, |
| 12 | maxTrainTime=1800, /* Set a 30-minute time limit */ |
| 13 | seed=911, |
| 14 | outputTables={names={ModelInfo='fraud_model_info', NObs='fraud_nobs', VariableSelection='fraud_var_select'}}; |
| 15 | RUN; |
| 16 | QUIT; |
3
Verify that the specified output tables were created.
Copied!
| 1 | PROC CAS; |
| 2 | TABLE.tableInfo TABLE={name='fraud_model_info'}; |
| 3 | TABLE.tableInfo TABLE={name='fraud_nobs'}; |
| 4 | TABLE.tableInfo TABLE={name='fraud_var_select'}; |
| 5 | RUN; |
| 6 | QUIT; |
Expected Result
The action should complete the training within the 30-minute 'maxTrainTime' limit. The use of 'distributeChains' should demonstrate the action's capability to handle distributed computation (in a real grid environment). The log should indicate that the training data was stored in memory. Three CAS tables ('fraud_model_info', 'fraud_nobs', 'fraud_var_select') should be created containing the results.