Shihang Zhu

Shihang Zhu

Milk Tea Buff

Undergraduate Student, School of Artificial Intelligence, Nanjing University

Language is Pale, while Vision is Alive.

Music atlas — multi-label style navigation

Music atlas — multi-label style navigation

Music page source and generated static assets for GitHub Pages.

Completion boundary

All 878 playlist IDs are in the catalog and graph. Public playlist IDs were reconciled against the local snapshot without missing or extra IDs. Advertised platform counts may fluctuate and are not used to invent extra songs.

This is not the completed 878-track musical interpretation requested by the owner. The prior 69-track sample also contains listening questions, rather than completed track-by-track listening. Its 112 candidate links retain that status. 789 tracks remain explicitly pending editorial interpretation.

The initial full-catalog experiment incorrectly let numerical feature clusters lead the musical composition. That was rejected by the owner. The non-style views keep the original six editorial exploration regions and original notes. The default style view arranges songs by sourced multi-label genre context. Unreviewed tracks are distributed by catalog relationships without assigning invented genres. Acoustic measurements and keyword statistics are secondary expandable information, with separate optional graph layers. No automatic measurement is called a finished listening interpretation.

Sources and evidence

  • _data/music.json: original playlist metadata; never overwritten by collection.
  • _data/music_atlas.json: stable-ID editorial notes, sources and candidate links.
  • _data/music_evidence.json: derived per-track measurements and provenance; no full lyrics, audio files, signed playback URLs or account credentials.
  • catalog.json: 878 tracks in original playlist order.
  • network.json: 878 nodes and 3009 typed edges at this snapshot. 289 same-album and 829 same-artist relations are identity facts. 720 excerpt acoustic links are measurements, 48 keyword links and 112 editorial links remain candidates.
  • Audio excerpts were measured for 594 tracks. The public API did not provide suitable audio for 284 tracks; no acoustic links are inferred for them.
  • Lyrics were available for 736 tracks. 114 other tracks were labelled instrumental by the platform, which is not independent proof that vocals are absent.
  • Sampling measures two approximately 20-second MP3 byte ranges at 25% and 65%. Byte positions only approximate chronological position for variable bitrate MP3s. Mono 16 kHz STFT descriptors cover centroid, bandwidth, flatness, low/high energy, frame-energy spread and spectral flux. These are not instruments, emotion, musical section labels, exact tempo, melody, harmony or full-track listening.
  • Acoustic edges require mutual four-nearest-neighbor membership and RMS distance at most .9 in seven transformed dimensions. Log transforms, 1/99-percentile winsorization, population z-scores and +/-3 clipping are recorded in the output.
  • Text links need at least two shared keyword themes, each with two distinct terms, and are capped at two links per track. They do not resolve negation or metaphor.
  • The default style view uses artist/release/track sources as explicitly scoped navigation context. Artist-level labels remain inferred for each track, not listening-confirmed classifications.
  • Raw API caches are ignored under local/music-atlas/evidence. Audio excerpts exist only in memory while measuring and are not saved or redistributed.

Rendering and design

Coordinates are precomputed during the Python build with deterministic spatial packing, identity springs and collision relaxation. No force iterations run on the browser’s main thread. ID sorting preserves layout after playlist reordering. PageRank is equal-weight and undirected over all typed relations, including candidates (damping .85); logarithmic radius 20–28 world units. It is not popularity or preference. Parallel relation types count as distinct edge records; PageRank adjacency treats an endpoint pair as one neighbor.

The graph draws at most 220 edges at a time and labels at most 28 desktop / 10 mobile songs. Labels are created on demand and avoid covers and region titles. Offscreen nodes are hidden. Overview artwork uses local 64px thumbnails, switching to 192px on zoom; covers/ and covers-small/ contain 718 unique archived covers. Thin perimeter colors are sampled from the outer 8% of artwork.

Every load defaults to static. Explicit optional motion moves region layers, rather than editing 878 node transforms every frame, and updates only drawn edges. Highlight freezes all layers; a second click on the same node clears it and resumes the previous motion setting. Escape and close clear selection. Background tabs, offscreen maps, list view and dragging pause motion. Reduced-motion defaults static and allows explicit opt-in. SVG geometry stays consistent with hit targets.

The intro preserves the title / Last Tide collage / right statement composition. The supplied Hugging Face collage is referenced directly without crop/color filters; clicking opens its 4096×2731 asset. Mobile places it below the title. Fonts are local with their SIL notices. The graph needs no third-party JavaScript runtime.

Multi-label style view

Default connection mode is 风格; default motion remains static. Every playlist ID has a style state. At this snapshot, 713 / 878 tracks have source-backed style labels; 165 remain 风格待确认. There are 156 observed genre labels. Unknown is a status, never a similarity label. No claim of full per-track listening.

Sources are stored in _data/music_style_sources.json (MusicBrainz), music_wiki_style_sources.json (artist infobox fields), music_yaogun_style_sources.json (Chinese Rock Database genre fields), and music_style_overrides.json (curated URLs and stable-ID song corrections). Country/language/year tags are rejected. Low-vote MusicBrainz outliers are excluded relative to the artist’s dominant genre votes. Known same-name mismatches, including Swiss Sadness versus the US project, are rejected. These remain navigation hints.

Genres retain parent labels, e.g. Midwest Emo → Emo → Rock, without inventing sibling tags such as Math Rock. The UI supports searchable AND multi-selection. One song has one node; its primary style sets its region, other direct genre labels create candidate bridges. Edges require a shared specific direct genre, exclude broad roots and pending status, and are capped at three per node. Scores combine inverse label frequency and label-set overlap, with a same-artist discount. Elliptical star groups and cover collision relaxation run only during the build.

To refresh artist sources before rebuilding:

python scripts/collect_music_styles.py
python scripts/collect_music_wiki_styles.py
python scripts/collect_music_yaogun_styles.py
python scripts/build_music_atlas.py

A track override can contain exclude, primary, and labels entries with id, url, title, and basis; these receive track-level provenance. Artist entries only receive release scope for exact albums name matches. Source caches remain ignored under local/music-atlas/.

Rebuild and collect

python -m pip install -r scripts/requirements-music-atlas.txt
python scripts/collect_music_evidence.py
python scripts/build_music_atlas.py
python scripts/build_music_atlas.py --check
python -m unittest discover -s scripts/tests

Collection uses public unauthenticated NetEase endpoints, bounded range requests and ffmpeg. It does not bypass unavailable audio or paid access. Cached unavailable results are explicit. ffmpeg is needed only for fresh audio measurement; checked-in website output has no Python runtime dependency. The external artwork archive is optional when the checked-in thumbnails exist.

python scripts/import_music_tracks.py exported-playlist.json
python scripts/import_music_tracks.py --playlist-id 8114560070

Imports preview by default; --apply writes and rebuilds. Stable IDs preserve editorial annotations. Duplicate or incomplete metadata aborts. New songs remain pending; they do not acquire guessed sound/genre/emotion interpretations. Backups live under ignored local/. Missing new artwork uses its provided URL and fallback.

python scripts/preview_music_atlas.py --port 4173

Local preview renders the actual include with relative-URL substitution. It is not a full Jekyll build; Ruby is unavailable and the full site build remains unverified.

Twelve data tests pass, including stable imports, PageRank mass, evidence boundaries, missing-lyrics handling, known-frequency signal measurement and full graph IDs. Actual browser checks cover 878 nodes, edge budget, search, mobile overflow, static default, highlight hold and repeated-click resume, with no page errors. Local reports: local/music-atlas/full-verification.json. Historical v5 timings measure the old browser layout and must not be presented as current timings.

Remaining: genuine track-by-track interpretation, stronger primary release/track sources, complete audio where legitimately available, listening-reviewed similarity links, full Jekyll verification and publication.