Classifying With or Without Labels

The classification pipeline can run in two modes, chosen per target from whether a QC flag column is available:

  • With labels β€” a flag is given, so ground-truth labels are built and the model’s performance is evaluated.

  • Without labels β€” no flag is available (unlabeled data), so the pipeline predicts every row but skips evaluation.

You do not need a dummy QC column for unlabeled data.

Option 1: with labels (default)

Give each target its QC flag and the flag values that mark positive and negative rows:

target_sets:
  - name: target_set_1
    variables:
      - name: temp
        flag: temp_qc
        pos_flag_values: [ 3, 4 ]
        neg_flag_values: [ 1, 2 ]

Rows are labelled from the flag, and the classify phase writes predictions plus the evaluation artefacts β€” the report, the model-scores file (see Performance Evaluation), and the metric plots.

Option 2: without labels

Omit flag (or set it to null / empty) for a target. The pipeline then keeps all rows, emits null flag/label columns, and skips label creation and performance evaluation:

target_sets:
  - name: target_set_1
    variables:
      - name: temp        # no `flag` β€” classify without evaluation

Predictions (predicted_label and score) are still produced; only the report, model-scores, and metric-plot files are skipped.

To force this for every target in a step regardless of flag, set the step-level override under the model step parameters:

step_param_set:
  steps:
    model:
      skip_evaluation: true   # omit to auto-detect per target from `flag`

Note

Only an absent or empty flag triggers the label-free path. A flag that names a column missing from the data is still treated as an error, so typos are not silently ignored.

SHAP values

SHAP explains predictions, not labels, so it is available in both modes and is controlled independently by calculate_shap (see SHAP Values):

step_param_set:
  steps:
    model:
      calculate_shap: true

In the label-free mode the label column of the SHAP output is null, while the per-feature SHAP contributions and the score are unaffected.