How to retrain a classifier in content explorer


Microsoft 365 compliance is now called Microsoft Purview and the solutions within the compliance area have been rebranded. For more information about Microsoft Purview, see the blog announcement.

A Microsoft 365 trainable classifier is a tool you can train to recognize various types of content by giving it samples to look at. Once trained, you can use it to identify items for application of Office sensitivity labels, communications compliance policies, and retention label policies.

This article shows you how to improve the performance of custom trainable classifiers by providing them additional feedback.

To learn more about the different types of classifiers, see Learn about trainable classifiers.

Watch this video for a quick summary of the tuning and retraining process. You'll still need to read this full article to get the details.


Pre-trained classifiers cannot be retrained.


To access classifiers in the Microsoft Purview compliance portal:

  • the Compliance admin role or Compliance Data Administrator is required to train a classifier

You'll need accounts with these permissions to use classifiers in these scenarios:

  • Retention label policy scenario: Record Management and Retention Management roles

Overall workflow


You provide feedback in content explorer for auto-apply retention label policies to Exchange items and uses the classifier as a condition. If you don't have a retention policy that auto-applies a retention label to Exchange items and uses a classifier as a condition, stop here.

As you use your classifiers, you may want to increase the precision of the classifications that they're making. You do this by evaluating the quality of the classifications made for items it has identified as being a match or not a match. After you make 30 evaluations for a classifier it takes that feedback and automatically retrains itself.

To understand more about the overall workflow of retraining a classifier, see Process flow for retraining a classifier.


A classifier must already be published and in use before it can be retrained.

How to retrain a classifier in content explorer

  1. Sign in to Microsoft Purview compliance portal with compliance admin or security admin role access and open Microsoft Purview compliance portal > Data classification > Content explorer.
  2. Under the Filter on labels, info types, or categories list, expand Trainable classifiers.


It can take up to eight days for aggregated items to appear under the trainable classifiers heading.

  1. Choose the trainable classifier you used in you auto-apply retention label policy. This is the trainable classifier you will give feedback on.


If an item has an entry in the Retention label column, it means that the item was classified as a match. If an item doesn't have an entry in the Retention label column, it means it was classified as a close match. You can improve the classifier precision the most by providing feedback on close match items.

  1. Choose an item and open it.


You can provide feedback on multiple items simultaneously by choosing them all and then choosing Improve classification in the command bar.

  1. Choose Provide feedback.
  2. In the Detailed feedback pane, if the item is a true positive, choose, Match. If the item is a false positive, that is it was incorrectly included in the category, choose Not a match.
  3. If there is another classifier that would be more appropriate for the item, you can choose it from the Suggest other trainable classifiers list. This will trigger the other classifier to evaluate the item.
  4. Choose Send feedback to send your evaluation of the match, not a match classifications and suggest other trainable classifiers. When you've provided 30 instances of feedback to a classifier, it will automatically retrain. Retraining can take from one to four hours. Classifiers can only be retrained twice per day.


This information goes to the classifier in your tenant, it does not go back to Microsoft.

  1. Open Trainable classifiers.
  2. The classifier that was used in your Communications compliance policy will appear under the Re-training heading.

classifier in retraining status.

  1. Once retraining completes, choose the classifier to open the retraining overview.

classifier retraining results overview.

  1. Review the recommended action, and the prediction comparisons of the retrained and currently published versions of the classifier.
  2. If you satisfied with the results of the retraining, choose Re-publish.
  3. If you are not satisfied with the results of the retraining, you can choose to provide additional feedback to the classifier in the Content Explorer interface and start another retraining cycle or do nothing in which case the currently published version of the classifier will continue to be used.

Details on republishing recommendations

Here is a little information on how we formulate the recommendation to re-publish a retrained classifier or suggest further retraining. This requires a little deeper understanding of how trainable classifiers work.

After a retrain, we evaluate the classifier's performance on both the items with feedback as well as any items originally used to train the classifier.

  • For built-in models, items used to train the classifier are the items used by Microsoft to build the model.
  • For custom models, items used in the original training the classifier are from the sites you had added for test and review.

We compare the performance numbers on both sets of items for the retrained and published classifier to provide a recommendation on whether there was improvement to republish.

See also