the quest for musical genres: do the experts and the wisdom of crowds agree?
DESCRIPTION
This paper presents some findings around musical genres. The main goal is to analyse whether there is any agreement between a group of experts and a community, when defining a set of genres and their relationships. For this purpose, three different experiments are conducted using two datasets: the MP3.com expert taxonomy, and last.fm tags at artist level. The experimental results show a clear agreement for some components of the taxonomy (Blues, HipHop), whilst in other cases (e.g. Rock) there is no correlations. Interestingly enough, the same results are found in the MIREX2007 results for audio genre classification task. Thus, showing the fact that a musical genre could have a multi–faceted definition; using expert based classifications, dynamic associations derived from the community driven annotations, and content–based analysis would improve genre classification, as well as other relevant MIR tasks such as music similarity or music recommendation.TRANSCRIPT
ISMIR / Philadelphia, US // September, 18th 2008
The Quest for Musical Genres:
Do the Experts and the Wisdom of Crowds Agree?
Mohamed Sordo, Òscar Celma, Martin Blech, Enric Guaus(Music Technology Group ~ UPF)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
motivation
taxonomy (controlled vocabulary)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
motivation
taxonomy (controlled vocabulary)
folksonomy (free text)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
motivation
taxonomy (controlled vocabulary)
VS. folksonomy (free text)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
expert-based
• taxonomy Mp3.com 2005
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
expert-based
• taxonomy 13 seed genres (components) 7 levels 711 genres
Rock Hip-Hop
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
community-based
• folksonomy last.fm
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
community-based
• folksonomy last.fm ~137K artists ~90K tags (after cleaning)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
outline
1) Mapping tags to genres
2) Computing similarity among genres
3) Agreement between experts and wisdom of crowds
4) Reconstructing the taxonomy from the folksonomy
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
outline
1) Mapping tags to genres
2) Computing similarity among genres
3) Agreement between experts and wisdom of crowds
4) Reconstructing the taxonomy from the folksonomy
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
• folksonomy ~ taxonomyJade (artist tags):
90s, illinois, new jack swing, rnb, r and b,
urban, ...
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
• folksonomy ~ taxonomyJade (artist tags):
90s, illinois, new jack swing, rnb, r and b,
urban, ...
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
• folksonomy ~ taxonomyJade (artist tags):
90s, illinois, new jack swing, rnb, r and b,
urban, ...
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
• folksonomy ~ taxonomyJade (artist tags):
R&B, New-Jack-Swing, Urban
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
1) mapping tags to genres
• folksonomy ~ taxonomyJade (artist tags):
R&B, New-Jack-Swing, Urban
(39% tags matched
with MP3.com genres)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
outline
1) Mapping tags to genres
2) Computing similarity among genres
3) Agreement between experts and wisdom of crowds
4) Reconstructing the taxonomy from the folksonomy
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
2) computing similarity among genres
• Taxonomy distance(Doo-Woop, Urban) = 3
Penalty when crossing components distance(Urban, Rock-Pop) = 7
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
2) computing similarity among genres
• Folksonomy LSA (SVD), 50 dim.
Cosine similarity sim(Urban, Doo-Wop) = 0.868 sim(Urban, Pop-Rock) = -0.145
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
outline
1) Mapping tags to genres
2) Computing similarity among genres
3) Agreement between experts and wisdom of crowds
4) Reconstructing the taxonomy from the folksonomy
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) Separate (taxonomy) genre components using (folksonomy) genre sim. intra-component similarity inter-component similarity
• 2) Correlation between (taxonomy) genre path distance and (folksonomy) genre sim. DistanceTAXONOMY(g1, g2) ~???~ SimFOLKSONOMY(g1, g2)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) intra-component similarity, using LSA
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) intra-component similarity, using LSA
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) intra-component similarity, using LSA
Alternative-Rap
Dirty-Rap
West-Coast
Hip-hop
Bass-Music
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) intra-component similarity, using LSA
Blu
es
Hip
-hop
Rock
/Pop
Ele
ctro
nic
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) inter-component similarity centroid for each component
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 1) inter-component similarity Clearly distinguishable from the rest
Hip-hop, Blues, Jazz
Relationships found Country ~ Bluegrass (~ Folk) R&B-Soul ~ Gospel/Spiritual Electronic/Dance ~ Vocal/Easy-Listening New-Age ~ World/Reggae (!)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 2) taxonomy genre distance vs. folksonomy genre sim.
West-Coast
Calypso
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 2) taxonomy genre distance vs. folksonomy genre sim. DistanceTAXONOMY(West-Coast, Calypso) = 8
West-Coast
Calypso
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 2) taxonomy genre distance vs. folksonomy genre sim. DistanceTAXONOMY(West-Coast, Calypso) = 7 SimFOLKSONOMY(West-Coast, Calypso) = 0.04
West-Coast
Calypso
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
agreement experts ~ wisdom of crowds
• 2) taxonomy genre distance vs. folksonomy genre sim.
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
outline
1) Mapping tags to genres
2) Computing similarity among genres
3) Agreement between experts and wisdom of crowds
4) Reconstructing the taxonomy from the folksonomy
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. Get genres at level n
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. For each genre at level n
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. Get all nodes at level n-1 (possible parents)
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. Compute cosine LSA similarity
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. Assign closest parent
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Select closest parent, using folk. genre sim. Compare with taxonomy parent
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Results
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Results
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
reconstruct taxonomy from folksonomy
• Results
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
...and also!
• MIREX 2007 results (Team: IMIRSEL-M2K SVM)
RAPHIPHOP 84.05%BLUES 77.68%EDANCE 77.68%JAZZ 72.53%COUNTRY 71.37%ROCKROLL 69.53%BAROQUE 65.81%METAL 61.11%ROMANTIC 52.79%CLASSICAL 33.33%
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
conclusions
• Consensus in some genres expert, community, and audio
• Discovery in terms of taxonomy/folksonomy coarse / fine grained static / dynamic
• Taxonomy adapts according to the folksonomy
• Do we need experts?
• Are some (wisdom-of-crowds) shepherds more experts than “THE” experts?
ISMIR / Philadelphia, US // September, 18th 2008 // òscar celma // MTG / UPF
future work
• Use more taxonomies and folksonomies
• Agreement measures
Uncovering affinity of artists to multiple genres from social behaviour data (Claudio Baccigalupo, Justin Donaldson, Enric Plaza)
ISMIR / Philadelphia, US // September, 18th 2008
THANKS!!!
Mohamed Sordo, Òscar Celma, Martin Blech, Enric Guaus(Music Technology Group ~ UPF)