Prediction Threshold

By default, a predicted label is 1 when the model’s positive-class score is at least 0.5, and 0 otherwise. This threshold is configurable.

Configuration

Set it under the model step parameters:

step_param_set:
  steps:
    model:
      predicted_label_threshold: 0.5    # default if omitted

• Default: 0.5, applied silently when the field is absent — existing configurations need no change.

• Range: any float in [0, 1]. A higher threshold makes positive predictions more conservative (fewer 1s); a lower threshold makes them more liberal.

• Scope: applies to every phase — validation, build/test, and classify — wherever a model converts scores into labels. It governs the predicted_label column of the user-facing prediction files and any threshold-dependent metrics in the reports.

How it is applied

The threshold is read from the configuration each time a model wrapper is constructed, and it is applied at prediction time as predicted_label = (score >= threshold).

There are three consequences worth understanding.

Warning

The threshold is not stored in the saved model file. Model files (*.joblib) contain only the trained estimator, not the threshold. When a saved model is loaded for the classify phase, the threshold comes from the classify configuration in effect at that time — not from whatever was used during training.

It is config-driven, not a runtime property.

Setting the threshold on a model object in code has no lasting effect if that object is later reconstructed from configuration (as happens, for example, when models are reloaded per target). Change the threshold by changing the configuration.

Keep it consistent across phases.

Because each phase reads its own configuration, using different thresholds in the training and classify configurations will produce labels at different operating points for the same model. Unless that is intended, set the same predicted_label_threshold in every phase’s configuration.

Relationship to model-scores files

The model-scores files described in Performance Evaluation are unaffected by this setting — they store raw score values, so you can evaluate performance across all thresholds regardless of which one is configured for label generation. The threshold only affects where the library draws the line when it must emit a concrete predicted_label.