Business Context
An engineering team is analyzing sensor data from industrial equipment to predict time-to-failure. For a critical component, they need not just the average prediction but the entire posterior distribution of predictions from the MCMC samples to perform advanced risk analysis and understand the full range of uncertainty.
Data Preparation
Generate a large training dataset (50k rows) and a scoring dataset (10k rows) simulating sensor readings. Train a model to predict 'time_to_failure'.
Copied!
| 1 | DATA mycas.sensor_train_large; |
| 2 | call streaminit(789); |
| 3 | DO i = 1 to 50000; |
| 4 | vibration = rand('NORMAL', 5, 1.5); |
| 5 | temperature = rand('NORMAL', 80, 10); |
| 6 | pressure = rand('NORMAL', 100, 5); |
| 7 | time_to_failure = 1000 - (temperature*5) - (vibration*20) - (pressure*1.5) + rand('NORMAL', 0, 50); |
| 8 | OUTPUT; |
| 9 | END; |
| 10 | RUN; |
| 11 | |
| 12 | DATA mycas.sensor_score_large; |
| 13 | call streaminit(101); |
| 14 | DO i = 1 to 10000; |
| 15 | vibration = rand('NORMAL', 5.5, 1.5); |
| 16 | temperature = rand('NORMAL', 85, 10); |
| 17 | pressure = rand('NORMAL', 102, 5); |
| 18 | OUTPUT; |
| 19 | END; |
| 20 | RUN; |
Étapes de réalisation
1
Train the BART model on the large sensor dataset.
Copied!
| 1 | PROC CAS; |
| 2 | bart.bartGauss / |
| 3 | TABLE='sensor_train_large', |
| 4 | inputs={{name='vibration'}, {name='temperature'}, {name='pressure'}}, |
| 5 | target='time_to_failure', |
| 6 | saveState={name='failure_model', replace=true}; |
| 7 | QUIT; |
2
Score the large scoring table, requesting all MCMC sample predictions by setting avgOnly=false. This will test performance and the ability to generate a wide table.
Copied!
| 1 | PROC CAS; |
| 2 | bart.bartScore / |
| 3 | TABLE='sensor_score_large', |
| 4 | restore='failure_model', |
| 5 | casOut={name='failure_posterior_preds', replace=true}, |
| 6 | avgOnly=false; |
| 7 | QUIT; |
3
Check the output table's columns to confirm the presence of individual sample predictions.
Copied!
| 1 | |
| 2 | PROC CAS; |
| 3 | TABLE.columnInfo / TABLE='failure_posterior_preds'; |
| 4 | QUIT; |
| 5 |
Expected Result
The action should execute successfully on the 10,000-row scoring table. The output table 'mycas.failure_posterior_preds' will be created. The column information should show the default 'Pred' column plus a series of columns for individual MCMC sample predictions (e.g., '_S_1', '_S_2', '_S_3', ...). The number of these columns depends on the default MCMC samples from the training step. The table should have 10,000 rows.