Scoring Recipe

When you score a multi-target regression model, DSS writes one prediction column per target. The output column names use the following format:

prediction_<target>

For example, a model with targets sales and margin outputs prediction_sales and prediction_margin.

Multi-target regression models can be deployed to the Flow as saved models and used in scoring recipes.

Available scoring engines

Multi-target regression scoring recipes can use the following scoring engines:

  • Python

  • Optimized (Java engine)

  • Spark

  • SQL (Snowflake with Java UDF)

The Python engine scores the model on the DSS server with the original Python model. The Optimized engine also scores on the DSS server, but uses the Java engine when the algorithm and preprocessing can be converted to this format.

For multi-target regression, Spark scoring and SQL (Snowflake with Java UDF) scoring have at least the same model compatibility requirements as the Optimized (Java engine). If a multi-target regression model is not compatible with the Optimized engine, it is not compatible with Spark or SQL (Snowflake with Java UDF) scoring either.

Regular SQL scoring is not supported. For more information about scoring engine compatibility, see Scoring engines.

Limitations

Multi-target regression has the following scoring limitations:

  • Only internal, unpartitioned models are supported for multi-target scoring recipes.

  • Prediction overrides are not supported.

  • Regular SQL scoring, PMML export, and Python export are not supported.