Importing similarity

Stars alone are a list. What makes the sky a map is similarity: who is listened to alongside whom. This guide imports that graph.

Run the MusicBrainz import first — the dataset is keyed by MusicBrainz artist credits, and resolving it needs the canon already in place.

Get the dataset

base=https://data.metabrainz.org/pub/musicbrainz/listenbrainz/labs/artist-credit-artist-credit-relations
dir=artist-credit-artist-credit-relations-02-20200813-178494
curl -O "$base/$dir/$dir.tar.bz2"

About 117 MB, roughly 7 million edges, CC0. Each line is one relationship:

{"id_0":1021,"name_0":"Ludwig van Beethoven","id_1":11285,"name_1":"Wolfgang Amadeus Mozart","score":1.0}

Run the import

lyrid import listenbrainz --dump ./artist-credit-artist-credit-relations-02-20200813-178494.tar.bz2

The whole dataset is read in one pass — about 7 million lines — and written in one transaction, so an interrupted run leaves the previous graph standing.

The import reports what it dropped rather than leaving the gaps invisible:

INFO lyrid::import::listenbrainz: resolving similarity against the canon credits=…
INFO lyrid::import::listenbrainz: dataset read; writing to PostgreSQL version=2020-08-13T12:41:07.693781+00:00 edges=… skipped_unknown=… skipped_weak=0 skipped_self=…
INFO lyrid::import::listenbrainz: similarity edges written rows=…
INFO lyrid::import::listenbrainz: artist prominence computed rows=…

• skipped_unknown — credits absent from your canon. A newer MusicBrainz dump than the similarity dataset will always produce some.
• skipped_self — both credits resolved to the same artist (“Nirvana” and “Nirvana feat. …”). An edge from a star to itself is not a route anywhere.
• skipped_weak — below --min-score, if you passed one.

Trimming the long tail

Most edges in the dataset are very faint — in a sample of half a million, only 14 scored above 0.5. Faint edges are not noise exactly, but they dominate the count without carrying much meaning:

lyrid import listenbrainz --dump ./relations.tar.bz2 --min-score 0.01

The threshold is recorded in the metric’s description, so a graph can always say how it was filtered.

Metrics

Similarity is not one fact but a family of them, so edges are grouped by metric:

SELECT key, description FROM similarity_metric;

Scores are comparable within a metric and never across metrics. Importing one metric replaces only its own edges, leaving the others standing. This is what later lets the same music be arranged by a deliberately different notion of closeness without discarding the first.

Brightness without listen counts

ListenBrainz publishes no per-artist listen counts as a dump — only a top-1000 API and a corpus licensed for non-commercial use only. So brightness is derived from the graph itself, in artist_prominence:

Column	Meaning
degree	How many edges the artist has
weight	Sum of those edges’ scores

weight is the more honest measure: one strong tie counts for more than many faint ones, which keeps a well-connected obscurity from outshining a genuinely central artist. The columns are named for what they measure, so nothing here can be mistaken for play counts.

Why a dataset from 2020

Current similarity exists at ListenBrainz only behind a live, rate-limited API. Covering three million artists would mean three million requests — exactly the failure that ADR 0002 was written to prevent, after a predecessor spent a month of hydration to collect 1,516 edges.

This dataset yields about four thousand times that, immediately. Musical similarity also ages slowly: Beethoven and Mozart have not drifted apart since 2020. What the vintage does cost is artists who rose afterwards, who arrive with no edges at all — which the design already has a place for, as the “dark matter” every map has at its margins.

Computing a fresh graph from the CC0 listen dumps is its own stage, and the schema is ready for it: it will land as a second metric beside this one.