Classifier configuration

A classifier-weight resource evaluates records that an indexer placed in the same candidate group. The indexer determines which records meet; the classifier determines the resulting classification.

Core settings

This is sample-customer-classifier from the Aurelia Utilities sample, shortened to four of its seven mappings:

{
  "type": "classifier-weight",
  "_id": "sample-customer-classifier",
  "description": "Decides whether two records in a bucket are the same person",
  "dataset": "sample-customer",
  "nonMatchThreshold": 0.55,
  "matchThreshold": 0.85,
  "bucketSizeReviewThreshold": 10,
  "defaultWeight": 1.0,
  "defaultOptions": ["AGGRESSIVE"],
  "mappings": [
    { "id": "tax-id",    "keys": ["taxId"],    "weight": 3.0, "options": ["AGGRESSIVE"], "object": false, "nullMismatch": false },
    { "id": "email",     "keys": ["email"],    "weight": 2.0, "options": ["AGGRESSIVE"], "object": false, "nullMismatch": false },
    { "id": "full-name", "keys": ["fullName"], "weight": 2.0, "options": ["AGGRESSIVE"], "object": false, "nullMismatch": false },
    { "id": "phone",     "keys": ["phone"],    "weight": 1.5, "options": ["AGGRESSIVE"], "object": false, "nullMismatch": false }
  ]
}

The remaining three are street at 1.0, and postcode and city at 0.5 each, for a total weight of 10.5.

Comparisons are derived from dataset column semantics and mapping settings. There is no comparer property with values such as LEVENSHTEIN or JARO_WINKLER in this resource model.

Text options include NONE, IGNORE_CASE, IGNORE_SPACES, ONLY_ALPHANUMERIC, IGNORE_ACCENTS, and AGGRESSIVE. Apply them according to the field’s business meaning; identifiers often need stricter normalization than names or addresses.

Comparison scores

The reported score summarizes comparisons according to the configured mappings and weights. Read both the result and its field-level evidence; a high score is not a probability that the records belong to one subject.

GET /api/golden/{entity}/duplicates/clusters/{clusterId}/evaluate returns those similarities so a reviewer can inspect the evidence. In the current customer sample, cleaning makes the two CRM phones for Ana comparable and their group is classified MATCH. The pair of different people named Javier Torres Melgar remains REVIEW, despite agreeing on name and postcode, because their tax identifiers differ. Inspect evaluate for the current per-field similarities and score; those values can move with the dataset, cleaner, indexer, and classifier definitions.

Mappings default to object: false, threshold: 50 for object comparison, and nullMismatch: false. A mapping without an explicit weight uses the classifier’s defaultWeight, initially 1.0.

Thresholds

reviewScoreThreshold accepts values above 0 and up to 1. It controls which cases appear in the default review queue, independently of their classification.

Configuration thresholds use a scale from 0 to 1. Reported cluster scores use integers from 0 to 100. A matchThreshold of 0.85 corresponds to a reported score of 85.

Multiply by 100 to go from one to the other, and state which one you mean when you write a runbook. A steward rule written with min: 0.9 does not select the clusters scoring 90 and above; it applies a threshold close to zero on the reported score scale. See Configure automatic stewardship.

Scores at or below nonMatchThreshold become NON_MATCH; scores at or above matchThreshold become MATCH; values between them become REVIEW.

bucketSizeReviewThreshold sends large groups to review as a safety control, whatever they scored.

The interval between the thresholds affects review volume. In the initial sample, thresholds of 0.55 and 0.85 yield one REVIEW group and 349 MATCH groups. Classification alone does not apply a merge.

Tune these settings with labeled representative examples. Do not copy the illustrative values above into production without measuring false matches, missed duplicates, and review volume for your data.

Related reading

• Configure an indexer
• Configure the merger
• Configure automatic stewardship
• Steward duplicate candidates
• Review duplicates manually, which walks these two clusters