Business Context
A financial services company needs to evaluate a fraud detection model across millions of transactions. To speed up the process and get segment-specific insights, they want to assess the model's performance grouped by transaction type (e.g., 'ONLINE', 'IN-STORE'). The dataset is partitioned by transaction date for efficiency.
About the Set : percentile
Precise calculation of percentiles and quantiles.
Discover all actions of percentile Data Preparation
Create a large, partitioned dataset simulating financial transactions. The table is partitioned by 'TXN_DATE' and includes a 'TXN_TYPE' for grouping. 'IS_FRAUD' is the target, and 'P_FRAUD' is the model score.
Copied!
| 1 | DATA casuser.transactions (partition=(TXN_DATE)); |
| 2 | FORMAT TXN_DATE yymmdd10.; |
| 3 | call streaminit(456); |
| 4 | array txn_types[2] $8 ('ONLINE', 'IN-STORE'); |
| 5 | DO TXN_DATE = '01NOV2025'd to '05NOV2025'd; |
| 6 | DO i = 1 to 200000; /* 1 million total records */ |
| 7 | TXN_ID = uuidgen(); |
| 8 | TXN_TYPE = txn_types[rand('integer', 1, 2)]; |
| 9 | IF TXN_TYPE = 'ONLINE' THEN base_prob = 0.05; ELSE base_prob = 0.01; |
| 10 | P_FRAUD = base_prob + rand('uniform')*0.1; |
| 11 | IS_FRAUD = rand('binomial', P_FRAUD, 1); |
| 12 | OUTPUT; |
| 13 | END; |
| 14 | END; |
| 15 | RUN; |
Étapes de réalisation
1
Run the assess action using the partitioned table. We specify `partition=true` for efficiency and use `groupBy` on 'TXN_TYPE' to get separate assessments for online and in-store transactions. We use the default 'ITERATIVE' method suitable for large data.
Copied!
| 1 | PROC CAS; |
| 2 | percentile.assess |
| 3 | TABLE={name='transactions', caslib='casuser', groupBy={'TXN_TYPE'}}, |
| 4 | response='IS_FRAUD', |
| 5 | inputs={{name='P_FRAUD'}}, |
| 6 | event='1', |
| 7 | partition=true, |
| 8 | includeFitStat=true, |
| 9 | fitStatOut={name='fraud_fitstat_by_type', caslib='casuser', replace=true}; |
| 10 | QUIT; |
2
Verify the output table contains separate statistics for each group.
Copied!
| 1 | |
| 2 | PROC CAS; |
| 3 | TABLE.fetch / TABLE={name='fraud_fitstat_by_type', caslib='casuser'}; |
| 4 | QUIT; |
| 5 |
Expected Result
The action should leverage the table's partitioning for efficient processing. The primary output, 'fraud_fitstat_by_type', must contain assessment statistics (like AUC and KS) for each level of 'TXN_TYPE'. The fetch result should show two rows in the fit statistics table, one for 'ONLINE' and one for 'IN-STORE', each with its own set of metrics.