Get started with trainable classifiers

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

Implementing a custom trainable classifier requires two steps:

1. Provide two sets of sample data (selected by humans).
   1. A set that contains only items that belong in the category.
   2. A set that contains only items that do not belong in the category.
2. Test the classifier's ability to detect matches.

This article explains how to create and test a custom classifier.

For more information about the different types of classifiers, see Learn about trainable classifiers.

Prerequisites

Licensing

For information on licensing, see

• Microsoft 365 Enterprise Plans
• Microsoft 365 Service Descriptions

Permissions

To use classifiers in the following scenarios, you need the following permissions:

Prepare for a custom trainable classifier

Before you create a custom trainable classifier, it's helpful to understand what's involved.

Overall workflow

For more information about the overall workflow of creating custom trainable classifiers, see the process flow for creating custom trainable classifiers.

Seed content

To ensure that your trainable classifier can independently and accurately identify that an item belongs to a particular category of content, you must present it with many samples of the type of content that is in the category. This feeding of samples to the trainable classifier is known as seeding. A human must select the seed content, and that content must include two sets of data: one set contains only items that strongly represent the content the classifier is designed to detect (positive samples) and a second set contains items that clearly don't belong (negative samples).

You need at least 50 positive samples (up to 500) and at least 150 negative samples (up to 1,500) to train a classifier. The more samples you provide, the more accurate the predictions the classifier makes will be. The trainable classifier processes up to the 2,000 most recently created samples (by file created date/time stamp).

How to create a trainable classifier

In preview: The following process automates the testing of trainable classifiers and shortens the creation workflow from 12 days to two days. In some cases, the process can take only a few hours.

1. Collect between 50 and 500 seed content items that strongly represent the data you want the classifier to positively identify as being in the category. For a list of supported file types, see Default crawled file name extensions and parsed file types in SharePoint Server.

2. Collect a second set of seed content (from 150 to 1,500 items) that represents data that don't belong in the category.

3. Place the positive and negative seed content in separate SharePoint folders. Each folder must be dedicated to holding only the seed content. Make note of the site, library, and folder URL for each set.

   Tip

   If you create a new SharePoint site and folder for your seed data, allow at least an hour for that location to be indexed before creating the trainable classifier that uses that seed data.

4. Sign in to the Microsoft Purview portal with either Compliance admin or Security admin role access and navigate to Data loss prevention > Data classification > Classifiers.

1. Select the Trainable classifiers tab.

2. Select Create trainable classifier.

3. Add the source of your positive examples: select the SharePoint site, library, and folder URL for the seed content that the classifier should detect and then choose Next.

4. Add the source of your negative examples: select the SharePoint site, library, and folder URL for the seed content that the classifier should ignore and then choose Next.

5. Review the settings and select Create trainable classifier.

6. Within 24 hours or less, the trainable classifier processes the seed data and builds a prediction model. The classifier status is In progress while it processes the seed data. When the classifier finishes processing the seed data, the status changes to Training is complete and items have been tested.

7. When training completes and items are (automatically) tested, publish the classifier by choosing Publish for use.

After you publish your classifier, it's available as a condition in Office auto-labeling with sensitivity labels, autoapply retention label policy based on a condition, and in Communication compliance.

Test your classifier

After the trainable classifier processes enough positive and negative samples to build a prediction model, test the predictions it makes. When you test the classifier, you verify whether its predictions are correct. After the classifier processes all the data, you can go through the results to verify whether each prediction is correct, incorrect, or uncertain. Microsoft uses this feedback in aggregate to improve the prediction model.

See also

• Learn about data loss prevention
• Sensitivity labels
• Sensitive information type entity definitions
• Document finger printing
• Learn about exact data match based sensitive information types