Settings

The “Settings” tab allows you to configure the multi-target regression task.

Target settings

The target settings define the set of numerical columns to predict.

The task must contain at least two targets, and all target columns must be numerical.

Changing the selected targets can require DSS to update feature roles and may reset settings that are no longer compatible with the new target list.

Train / test split

Multi-target regression uses the same train/test settings as single-target prediction models. The default setup uses the standard train/test split. K-fold cross-test and explicit train/test extracts are supported. For more information, see Settings: Train / Test set.

Metrics

Multi-target regression uses regression metrics. The default optimization metric is R2.

DSS computes metrics for each target and also computes global metrics by averaging the per-target metric values. For hyperparameter search and model comparisons in the Result screen, the optimized metric is the global version of the selected metric.

Custom metrics are supported in the model analysis. By default, DSS applies the custom metric function independently to each target, then uses the arithmetic mean across targets. You can provide a custom target aggregation function to override this default aggregation.

Algorithms

The available algorithms are Random Forest, Ridge Regression, and Neural Network from scikit-learn. The default setup enables Random Forest with hyperparameter search.

Spark MLLib, Vertica, deep learning, custom algorithms, and ensemble models are not supported for multi-target regression.

For algorithms compatible with Java scoring, multi-target regression models can also use compatible Java-based scoring engines. For more information, see Scoring Recipe.

Features handling

Multi-target regression uses the same feature handling settings as classical prediction models. For more information, see Features handling.

The following differences apply:

  • Target encoding for categorical input features is not supported.

  • Feature reduction is limited to no reduction, PCA, or ICA.

  • Monotonic constraints are not supported.

Runtime and backend

Multi-target regression training uses the Python in-memory backend.

Runtime environment settings are identical as for other Python in-memory prediction tasks. Spark (MLLib) training backend is not supported.

Sample weights

The purpose of a “sample weight” variable is to specify the relative importance of each row of the dataset, both for the training algorithm and for the different evaluation metrics. The sample weight column must be different from all target columns.

Limitations

The following limitations apply:

  • Target columns must be numerical.

  • The task must contain at least two targets.

  • Partitioned models are not supported.

  • Model ensembling is not supported.

  • Target encoding for categorical input features is not supported.

  • Feature reduction is limited to no reduction, PCA, or ICA.

  • Monotonic constraints are not supported.