Entity resolution: Discogs masters against MusicBrainz release groups

Three methods score the same blocked candidate pairs and are judged on the same test fold against labelled ground truth. Unlinked MusicBrainz records are unlabelled, not known non-matches: they count toward coverage and unverified accepts, never precision, recall or F1. Every figure here is read from the JSON files the dbt project exports into data/; the page types no number of its own.

Decision summary

Loading the artifact-backed comparison…

Methods on the test fold

methodpair completenessprecisionrecall (labelled)recall (overall)F1coverageunverified acceptsreview queueECE (decisions)

Source: fct_eval_metrics.json. Precision, recall and F1 are over labelled A records; coverage counts every test-fold A record; unverified accepts are auto-accepts on unlabelled A records, counted by coverage and not by the accuracy metrics.

Reliability: top-candidate probability against observed correctness

exact_v1rules_v1learned_v1diagonal: perfect calibration

Source: fct_calibration.json, decision level. Point size grows with the bin count.

Review cost against the accept threshold

exact_v1rules_v1learned_v1marker: chosen threshold

Source: fct_review_queue.json. Total cost in review-row units at the cost ratio selected: .

What a review budget buys

exact_v1rules_v1learned_v1

Source: fct_review_floor.json. Each point is one review floor; the horizontal axis is the queue as a share of the test fold, the vertical axis is recall if every queued row were resolved.

Blocking

keypairspair completeness

Source: fct_blocking.json.

Sources

sourcesidedumpdatelicence

Source: dim_source.json. Licence texts are in docs/DATA_SOURCES.md of the repository.