Skip to content

Purview Classifiers and Sensitive Information Types

Atlanta, USA (United States of America)

GitHub Cloud2BR OSS (open-source software) - Learning Hub

Last updated: 2026-07-27

References

Back to the documentation hub

Classifiers are how Purview recognizes sensitive content. Accurate classification is the foundation for labels, data loss prevention (DLP), retention, and insider risk, so invest in tuning it before you enforce policies that depend on it.

Choose the right classifier

Method Best for Primary check
Sensitive information type (SIT) Structured identifiers such as regulated numbers Pattern, keywords, confidence, and proximity
Exact data match (EDM) Matching against your own sensitive data set Data refresh, hashing, and schema accuracy
Trainable classifier Categories defined by example, not pattern Quality and volume of training samples
Named entity and credential SITs Common entities and leaked secrets False-positive tolerance and scope

Tune before you enforce

  1. Define the data category, the owner, and the policy that will consume it.
  2. Choose the classifier method that fits the data shape and accuracy need.
  3. Test against representative true-positive and benign content.
  4. Adjust confidence level, supporting evidence, and count thresholds.
  5. Validate results in content explorer and activity explorer before enforcement.

Operational controls

  • Record each classifier's purpose, owner, confidence, and dependent policies.
  • Re-test classifiers after data, schema, or policy changes.
  • Restrict who can view classified content and export classification results.
  • Prefer a small, well-understood set of classifiers over many noisy ones.

Verify and operate

  • Measure precision and recall against known true-positive and benign content, not just spot checks.
  • Use content explorer to confirm where a classifier matches before any policy consumes it.
  • Track classifier changes and re-test dependent labels, DLP (data loss prevention), and retention policies.
  • Retire or merge overlapping classifiers that add noise without adding coverage.

Business example

A team needs to protect a proprietary product-code format. A custom sensitive information type with a validating function and supporting keywords is tested against sample documents, tuned to reduce false positives, and only then linked to a sensitivity label and a DLP (data loss prevention) policy in simulation mode.