3.9.6 Scoring Models

ML_SCORE scores a model by generating predictions using the feature columns in a labeled dataset as input and comparing the predictions to ground truth values in the target column of the labeled dataset.

The dataset used with ML_SCORE should have the same feature columns as the dataset used to train the model but the data sample should be different from the data used to train the model; for example, you might reserve 20 to 30 percent of a labeled dataset for scoring.

ML_SCORE returns a computed metric indicating the quality of the model. A value of None is reported if a score for the specified or default metric cannot be computed. If an invalid metric is specified, the following error message is reported: Invalid data for the metric. Score could not be computed.

Models with a low score can be expected to perform poorly, producing predictions and explanations that cannot be relied upon. A low score typically indicates that the provided feature columns are not a good predictor of the target values. In this case, consider adding more rows or more informative features to the training dataset.

You can also run ML_SCORE on the training dataset and a labeled test dataset and compare results to ensure that the test dataset is representative of the training dataset. A high score on a training dataset and low score on a test dataset indicates that the test data set is not representative of the training dataset. In this case, consider adding rows to the training dataset that better represent the test dataset.

HeatWave ML supports a variety of scoring metrics to help you understand how your model performs across a series of benchmarks. For ML_SCORE parameter descriptions and supported metrics, see Section 3.10.8, “ML_SCORE”.

Before running ML_SCORE, ensure that the model you want to use is loaded; for example: 

CALL sys.ML_MODEL_LOAD(@census_model, NULL);

For information about loading models, see Section 3.9.3, “Loading Models”.

The following example runs ML_SCORE to compute model quality using the balanced_accuracy metric: 

CALL sys.ML_SCORE('heatwaveml_bench.census_validate', 'revenue', @census_model, 
'balanced_accuracy', @score);

where:

  • heatwaveml_bench.census_validate is the fully qualified name of the validation dataset table (schema_name.table_name).

  • revenue is the name of the target column containing ground truth values.

  • @census_model is the session variable that contains the model handle.

  • balanced_accuracy is the scoring metric. For other supported scoring metrics, see Section 3.10.8, “ML_SCORE”.

  • @score is the user-defined session variable that stores the computed score. The ML_SCORE routine populates the variable. User variables are written as @var_name. The examples in this guide use @score as the variable name. Any valid name for a user-defined variable is permitted (e.g., @my_score).

To retrieve the computed score, query the @score session variable. 

SELECT @score;
+--------------------+
| @score             |
+--------------------+
| 0.8188666105270386 |
+--------------------+