A field's role controls how it's used in model building—for example, whether a field is an input or target (the thing being predicted).

The following roles are available:

Input. The field is used as an input to machine learning (a predictor field).

Target. The field is used as an output or target for machine learning (one of the fields that the model will try to predict).

Both. The field is used as both an input and an output by the Apriori node. All other modeling nodes will ignore the field.

None. The field is ignored by machine learning. Fields whose measurement level is set to Typeless are automatically set to None in the Role column.

Partition. Indicates a field used to partition the data into separate samples for training, testing, and (optional) validation purposes. The field must be an instantiated set type with two or three possible values (as defined in the advanced settings by clicking the gear icon). The first value represents the training sample, the second represents the testing sample, and the third (if present) represents the validation sample. Any additional values are ignored, and flag fields can't be used. Note that to use the partition in an analysis, partitioning must be enabled in the node settings of the appropriate model-building or analysis node. Records with null values for the partition field are excluded from the analysis when partitioning is enabled. If you defined multiple partition fields in the flow, you must specify a single partition field in the node settings for each applicable modeling node. If a suitable field doesn't already exist in your data, you can create one using a Partition node or Derive node. See Partition node for more information.

Split. (Nominal, ordinal, and flag fields only.) Specifies that a model is built for each possible value of the field.

Frequency. (Numeric fields only.) Setting this role enables the field value to be used as a frequency weighting factor for the record. This feature is supported by C&R Tree, CHAID, QUEST, and Linear nodes only; all other nodes ignore this role. Frequency weighting is enabled by means of the Use frequency weight option in the node settings of those modeling nodes that support the feature.

Record ID. The field is used as the unique record identifier. This feature is ignored by most nodes; however, it's supported by Linear models.