Skip to content

The ML: Score step scores a data table with a model trained by ML: Train Model and writes the data with one appended prediction column. Because the model arrives as an ordinary table, you can train once and score many tables, or re-score whenever new data lands. The step runs as a managed PlaidCloud job (like the LLM step).

model artifactfrom ML Trainnew rowsfeatures onlyScoreapply modelrows + predictionnew column
ML Score applies a saved model to new rows that carry only features, appending a prediction column to each.
  • Data Table — the table to score. It must contain every feature column the model was trained on.
  • Model Table — a model table written by an ML: Train Model step. The step fails if the table is empty, or if any part of the stored model is missing.
  • Output Table — where the scored data is written.
  • Prediction Column — the name of the appended column. Leave blank to use prediction. The step fails if the name already exists in the data table.

The output table contains every column of the data table plus the prediction column: text class labels for classifiers, numeric values for regressors. Feature values are prepared exactly as they were at training time, and categorical values the model never saw during training are ignored rather than causing an error.

The model table must have the shape an ML: Train Model step writes — the full set of model columns, a known algorithm, and every part of the stored model present. The model itself is loaded under a restriction that rejects anything outside the standard scientific stack, so a hand-crafted model row cannot run arbitrary code.

The Scoring Options section has these fields. Converted KNIME workflows use them too.

  • Class Probability Columns (probability_column_template) — for a classifier, append one probability column per class, named by this template with {class} replaced. Probabilities before the prediction (probabilities_first) puts them before the prediction column, and Order classes as they first occur in training (classes_in_domain_order) orders them by the order the classes first occur in the training data instead of sorted.
  • Confidence Column (confidence_column) — for a classifier, append the winning class’s probability under this name.
  • Variance Column (variance_column) — for a forest regressor, append the variance of the trees’ predictions under this name.
  • Model Count Column (model_count_column) — for a tree model, append the number of trees under this name.
  • Appended Column Order (column_order) — the order of the appended columns. Blank: prediction, probabilities, confidence, variance, then model count.
  • Write Instead (report) — write one of the model’s own tables instead of the scored data, such as variable_importance, coefficients, statistics, pmml (the model as a PMML document), centers, cluster_counts, or association_rules.
  • Model Format (model_format) — pmml scores with a PMML document: the model table holds the document, or the model of a learner that has a PMML form.
  • Algorithm Settings (algorithm_settings) — settings of the scoring or the report particular to the model’s algorithm, such as the number of components a pca model projects to.
  • Third Table (source_c) — a third table that some algorithms read, such as the random forest a knn_distance model uses.

An AI assistant or the REST API can also set report to shap_values.

  • In-memory working set — the data table is read into memory, so the step targets roughly one million rows per run.
  • Converted Alteryx workflows — a converted Predict tool pins the prediction column to the target variable’s name with a _predicted suffix.