Skip to content

Click in-app to access the full platform documentation for your version of DataRobot.

Restore features removed by reduction

Availability information

Restoring features that were pruned after the feature derivation process, available as a public preview feature, is off by default. Contact your DataRobot representative or administrator for information on enabling the feature.

Feature flag: Enable Restoration of Time Series Reduced Features

In any time series project, DataRobot generates derived features based on the window settings at project start. DataRobot then runs a feature reduction algorithm, removing features it detects as low impact. Sometimes, however, the algorithm may remove some important features during the feature reduction process—features that you want included in the generated feature lists or evaluated for feature impact. Some examples of this are certain calendar-derived features or a particular numeric statistic of a financial variable. After EDA2 completes, you can add these features back into your available derived modeling data.

Note

Even if you disable supervised reduction in advanced options, DataRobot may still remove features based on extractor priority. These features can also be restored with the restoration process.

Identify removed features

The easiest way to determine whether features were removed in the feature reduction process is to review the feature derivation log after EDA2 completes.

Depending on the dataset size, it is likely you need to download the log. This is because the reduction process runs last (is at the end of the file) and may be truncated from the preview.

Restore pruned features

The following describes how to restore removed features (identified from the derivation log) to the modeling dataset. You can use this option repeatedly, until you have restored all features or have reached the maximum supported features, which may be constrained by data ingest limits.

  1. On the Data > Derived Modeling Data tab, select Restore pruned features from the menu:

  2. In the Restore pruned features window, begin typing to select features for restoration. DataRobot indicates the number of features that can be added back.

  3. Click Add features when all desired features are listed. DataRobot reports progress:

    And then success:

Note

Feature restoration does not change the feature lists created during EDA2. To use the restored features for modeling, create new feature lists.

Verify restored features

There are two methods for verifying that features were restored. In these examples, the process was run twice, adding different features each time:

  • Sort features by index number on the Data page. Restored features have the highest index (most recently added), so they are shown first.

  • Restoration is also recorded to the derivation log. Download the log and check for restored features. This sample log shows results of more than one restoration process:

Create new feature lists

When features are restored, they are not added into existing feature lists. To use the new features as part of your modeling dataset you must create new feature list(s) that incorporates them. For example:

  1. From the Derived Modeling Data tab, select the best performing feature list. Check the Feature Name box to select all features in that list.

  2. Change to the All Time Series Features list (selections from the previous action are preserved).

  3. Select the restored features you would like to add.

  4. Click Create feature list to add the new list.

Once one or more new lists are created that contain the restored features, build models with them (individually or by rerunning Autopilot). Compare model performance between lists to see if there is value in including the restored features as part of the model to use for making predictions.


Updated October 28, 2021
Back to top