Skip to main content

Explanations and training

Explanations​

Figure 1. Explanations and training runs.

Global feature importance as reported for the latest version, and how each feature moved since the version before.

Top features lists the features by weight. The header names the version, the version its ranks are compared against and the method, as your pipeline or MLflow reported it (for example method: gain). Each feature shows:

  • Its rank and weight.
  • A rank change badge, such as Up 1 or Down 1, since the previous version.
  • Its previous rank, for example was #3.

The first version has no earlier version to compare, so no rank changes are shown.

Training runs​

Every training run for the model, newest first: Run, Version, Dataset version, Status, Metrics (for example f1 0.840 · auc 0.920) and Started.

A failed retraining run stays in the list with its error next to the failed status, for example failed OutOfMemoryError. A failed run can also raise an alert (see Alerts).

Explain a prediction​

Figure 2. Explain a prediction.

Look up one scored row to see its features, its prediction and the attributions behind it.

  1. On the model page, go to Explain a prediction.
  2. Enter the Row id your pipeline sent with the row.
  3. Select Look up.

You see:

PanelShows
PredictionThe Row, when it was Scored, the Prediction, the Predicted label, the actual Label and the row's protected attributes.
FeaturesThe inputs the model scored.
Why this predictionEach feature's contribution to this prediction, as your pipeline sent it: its size, and whether it raised or lowered the prediction.