Data classification in Microsoft Purview

Data classification in Microsoft Purview is how the platform recognizes which content is sensitive, so that labels, DLP, retention and insider risk policies have something to act on. Learn contrasts three ways of identifying items: manually by users, by automated pattern matching (sensitive information types), and by machine learning (trainable classifiers). The decision rule: if the content has a recognizable pattern, use a sensitive information type; if it is an exact known value, use exact data match; if it is a form that varies, use document fingerprinting; if it is recognizable only by meaning, use a trainable classifier.

The classifier ladder

Pick the simplest tool that fits and move down only when it fails.

Person names, physical addresses and medical terms are covered by named entity SITs, which use dictionaries and patterns, appear by default and cannot be edited or copied.

How a sensitive information type decides

A SIT is a pattern made of a primary element (a regular expression with or without a checksum, a keyword list or dictionary, or a function), optional supporting elements that corroborate it, a confidence level, and a proximity (how many characters may separate evidence from the primary element). Learn’s example uses 250 characters, and the supporting element must lie entirely inside the window to count.

Confidence reflects how much supporting evidence was found. There are three levels, and Learn gives their numeric values:

• Low (65): fewest false negatives, most false positives. Returns low, medium and high matches.
• Medium (75): average of both.
• High (85): fewest false positives, most false negatives. Returns only high matches.

Learn advises high-confidence patterns with low match counts (five to ten) and low-confidence patterns with higher counts (twenty or more). This is the tuning lever: a DLP rule that fires too often needs higher confidence or a higher count; one that misses real data needs the opposite. Policies using older numeric “accuracy” values are mapped to the three levels (76 to 100 high, 66 to 75 medium, 65 or below low).

Trainable classifiers

Learn describes trainable classifiers as suited to content that cannot be identified manually or by pattern, and notes that a classifier learns from hundreds of examples. Two kinds exist: pretrained classifiers from Microsoft (status Ready to use) and custom classifiers you train with a set of positive examples and a set of negative ones, then confirm the predicted true and false positives and negatives before publishing. Points the exam likes: custom classifiers are English-only, classifiers only work on items that are not encrypted, and a published custom classifier cannot be retrained, so you remove it and start again with larger samples. Classifiers can be a condition for auto-applied sensitivity labels, auto-applied retention labels, DLP and, with Microsoft-provided classifiers only, communication compliance.

Worked example

Fabrikam wants to protect (1) employee bank account numbers, (2) a customer table of 2 million rows, and (3) signed supplier contracts.

1. Bank account numbers follow a format. A built-in SIT for the country exists. Medium confidence with a count of 10 or more cuts noise in a DLP rule.
2. The customer table must match real customers, not any nine-digit number. Use an EDM based SIT built from the table, which Learn describes as detecting exact values in a database of sensitive information that can be refreshed daily and hold up to 100 million rows.
3. Contracts have no fixed pattern. Train a custom classifier with contracts as positive samples and unrelated documents as negatives, test it, publish it, and then use it as the condition of an auto-labeling policy for sensitivity labels.

Afterward, Data explorer (the successor to Content explorer, which Learn now labels classic) and Activity explorer show where the classified content lives and what happens to it, which is how classification is checked before enforcement. The Classifiers page, in preview as a combined experience at the time of writing, lists SITs, EDM, trainable classifiers and fingerprints in one place.

Which exams test this

For how classification feeds enforcement, see DLP explained for SC-401.

How the questions are built

• “A pattern, such as a card number” is a SIT. “Exact values from our database” is EDM. “Variations of a template” is fingerprinting. “No pattern, identified by what it is” is a trainable classifier.
• “Too many false positives” means raise confidence or the instance count. “Missing real matches” means lower them.
• “Retrain a published classifier” is a trap: it is not supported.
• “Encrypted files are not classified” is a classifier limit stated by Learn.
• “Where is sensitive data and what is happening to it” points to Data explorer or Content explorer (items) or Activity explorer (events), not to a policy.

Go deeper

• Learn about sensitive information types (Microsoft Learn)
• Learn about trainable classifiers (Microsoft Learn)
• Microsoft Purview classifiers overview (Microsoft Learn)
• Related topics: sensitivity labels explained, DLP explained, retention labels vs policies