A Collaborative Music Knowledge Graph Dataset for Artist Recommendation Based on Open Data Sources

This deposit contains the input resources and generated dataset for a collaborative music knowledge graph built from open music data sources.

The dataset combines MusicBrainz metadata with Last.fm listening data to represent users, artists, genres, labels, and geographic areas, together with semantic and listening relationships such as user-artist preferences, artist genres, artist labels, artist areas, and artist membership relations.

The uploaded files include two main folders:

resources/

Contains the input data used by the processing pipeline:

resources/lastfm: Last.fm user listening data, including user-artist-playcount records.

resources/mbdump: selected MusicBrainz dump tables required by the dataset generation process.

dataset/

Contains the generated outputs:

dataset/db/staging.db: SQLite database produced from the MusicBrainz and Last.fm input files.

dataset/graph/musicgraph.graphml: full generated music knowledge graph in GraphML format.

dataset/*.csv: CSV exports of graph nodes and edges, grouped by node or relation type.

The CSV files are generated from the GraphML graph and then cleaned automatically by removing auxiliary export columns such as id, type, and relation. The resulting CSV files are intended for easier import into graph databases, data analysis tools, or downstream recommender-system experiments.

The generated dataset was produced with the project Makefile using the workflow:

make generate_dataset_with_csv

Default sampling parameters:- SAMPLE_SIZE = 0.05- MIN_USER_PLAYS = 20000- MIN_USER_ARTISTS = 40- ARTIST_TOP_N = 1- TEST_DATA_SIZE = 0.5

Source data references:- MusicBrainz database dumps: https://musicbrainz.org/doc/MusicBrainz_Database/Download- Last.fm 360K dataset: https://ocelma.net/MusicRecommendationDataset/lastfm-360K.html

User identifiers from the Last.fm data are stored as hashed user IDs as provided by the original dataset.

CSV file format

The dataset/csv folder contains cleaned CSV exports of the generated music knowledge graph. Files are split into node files and edge files. The CSV files are plain UTF-8 comma-separated files with a header row.

Node CSV files:

nodes_user.csv: user nodes. Columns: user_sha.

nodes_artist.csv: artist nodes. Columns: mbid, artist_name.

nodes_genre.csv: genre nodes. Columns: genre_name.

nodes_label.csv: label nodes. Columns: mbid, label_name.

nodes_area.csv: geographic area nodes. Columns: area_name.

Edge CSV files:

edges_favours_artist.csv: listening/preference relations from users to artists. Columns: user_sha, artist_mbid, plays.

edges_from_area.csv: relations from artists to geographic areas. Columns: artist_mbid, area_name.

edges_hasGenre.csv: relations from artists to genres. Columns: artist_mbid, genre_name.

edges_hasLabel.csv: relations from artists to labels. Columns: artist_mbid, label_mbid.

edges_member_of.csv: artist membership relations, for example solo artists or members linked to groups. Columns: from_artist_mbid,to_artist_mbid.

In all edge files, source and target contain the identifiers of the connected graph nodes. For artist and label nodes these identifiers correspond to MusicBrainz MBIDs when available; for users they correspond to hashed Last.fm user identifiers; for genres and areas they correspond to their exported graph node names.

Stats

Current CSV contents:

nodes_user.csv: 2,226 rows

nodes_artist.csv: 16,962 rows

nodes_genre.csv: 620 rows

nodes_label.csv: 2,549 rows

nodes_area.csv: 542 rows

edges_favours_artist.csv: 109,828 rows

edges_from_area.csv: 16,962 rows

edges_hasGenre.csv: 26,197 rows

edges_hasLabel.csv: 40,092 rows

edges_member_of.csv: 1,535 rows