fix: ListenBrainz similarity is stored relative to the seed's best, and the candidate arms keep tier order (#6177)
release / govulncheck (push) Successful in 15s
release / web (push) Successful in 1m23s
release / go (push) Successful in 1m36s
release / integration (push) Successful in 5m4s
release / android (push) Successful in 5m5s
release / Build signed APK (releases and dev) (push) Successful in 5m18s
release / Attach APK to the Release (tag releases only) (push) Skipped
release / Build + push container image (push) Successful in 30s
release / Verify release artifacts (tag releases only) (push) Skipped
release / govulncheck (push) Successful in 15s
release / web (push) Successful in 1m23s
release / go (push) Successful in 1m36s
release / integration (push) Successful in 5m4s
release / android (push) Successful in 5m5s
release / Build signed APK (releases and dev) (push) Successful in 5m18s
release / Attach APK to the Release (tag releases only) (push) Skipped
release / Build + push container image (push) Successful in 30s
release / Verify release artifacts (tag releases only) (push) Skipped
LB's similarity scores are counts (16 to 1445 for recordings, to 10,875 for artists), and they reached a scorer that reads similarity as 0 to 1. One edge added 64 to 5,800 points under Songs-like, so likes, skips and recency stopped mattering for any candidate with an edge. - ResolveListenBrainzTrackEdges divides each edge by the seed's best score in its whole cached answer. The cache keeps the raw answer. - The worker stores matched and unmatched artist rows over the seed's top artist. - Migration 0096 rescales the rows already stored. - The candidate arms hold sim_score bands so tier 1 stays above tier 2: lb_similar 0.5 + 0.5 x edge, tag_overlap halved, likes_overlap 0.3. Otherwise a seed's weaker LB edges fell below a shared genre tag. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -199,6 +199,10 @@ func (w *Worker) upsertArtistSimilar(ctx context.Context, q *dbq.Queries, artist
|
||||
return
|
||||
}
|
||||
sort.Slice(results, func(i, j int) bool { return results[i].Score > results[j].Score })
|
||||
// Scores are stored relative to the seed's closest artist, as track edges
|
||||
// are (#6177): LB's raw scores are counts (16 to 10,875 on the operator's
|
||||
// library), and both the candidate pool and the suggestions read 0 to 1.
|
||||
best := results[0].Score
|
||||
|
||||
mbids := make([]string, 0, len(results))
|
||||
for _, r := range results {
|
||||
@@ -230,7 +234,7 @@ func (w *Worker) upsertArtistSimilar(ctx context.Context, q *dbq.Queries, artist
|
||||
continue
|
||||
}
|
||||
if uerr := q.UpsertArtistSimilarity(ctx, dbq.UpsertArtistSimilarityParams{
|
||||
ArtistAID: artistAID, ArtistBID: localID, Score: r.Score,
|
||||
ArtistAID: artistAID, ArtistBID: localID, Score: relativeScore(r.Score, best),
|
||||
}); uerr != nil {
|
||||
w.logger.Warn("similarity: UpsertArtistSimilarity", "err", uerr)
|
||||
continue
|
||||
@@ -257,7 +261,7 @@ func (w *Worker) upsertArtistSimilar(ctx context.Context, q *dbq.Queries, artist
|
||||
SeedArtistID: artistAID,
|
||||
CandidateMbid: r.MBID,
|
||||
CandidateName: r.Name,
|
||||
Score: r.Score,
|
||||
Score: relativeScore(r.Score, best),
|
||||
Source: "listenbrainz",
|
||||
}); uerr != nil {
|
||||
w.logger.Warn("similarity: UpsertArtistSimilarityUnmatched", "err", uerr)
|
||||
@@ -266,3 +270,12 @@ func (w *Worker) upsertArtistSimilar(ctx context.Context, q *dbq.Queries, artist
|
||||
takenUnmatched++
|
||||
}
|
||||
}
|
||||
|
||||
// relativeScore is score over the seed's best, in [0, 1]; 0 when the seed has
|
||||
// no positive score to measure against.
|
||||
func relativeScore(score, best float64) float64 {
|
||||
if best <= 0 {
|
||||
return 0
|
||||
}
|
||||
return score / best
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user