Page 1
Exploiting Recommended Usage Metadata: Exploratory Analyses
Xiao Hu, J. Stephen Downie, Andreas Ehmann
THE ANDREW W. MELLON FOUNDATION
The International Music Information Retrieval Systems Evaluation Lab
(IMIRSEL)University of Illinois at Urbana-
Champaign
Page 2
Motivation
Human Use of Music Information Retrieval Systems (HUMIRS) project
to identify: Standardized MIR evaluation tasksQuery documents
from real world users’ behaviors
User generated usage metadata
Page 3
Usage Metadata
Music Customer reviews on www.epinions.com
Each review is associated with one recommended usage
Page 4
Usage Categories
Driving Waking up
Hanging With Friends Going to Sleep
Listening Cleaning the House
Romancing At Work
Reading or Studying With Family
Getting ready to go out
Sleeping
Exercising
Prepared by epinions.com editors
Page 5
Album Metadata
Each review is for an album
album title
Page 6
Research Questions
Q1: What are the relationships between usages and music genres?
Q2: What are the relationships between usages and music artists?
Q3: How are the usages related to each other?
Page 7
Data Facts
Number of Usage Categories 11Reviews in Each Usage Categories
180
Total Number of Reviews 1,980Number of Genres 12
Number of Artists 897
Number of Album titles 1,372
Page 8
Genre and Usage (1)
Genres:
Blues Heavy Metal
Classical International
Country Jazz Instrument
Electronic Pop Vocal
Gospel R&B
Hardcore/Punk
Rock & Pop
Page 9
Genre and Usage (2) Dependency analysis:
Pearson’s chi-square dependency test on each pair of genre and usage (p < 0.001)
Genre Usage Pearson’s χ2
Classical Listening 37.613
Country Cleaning the House
70.782
Electronic Going to Sleep 29.127
Hard Core / Punk
Waking Up 12.536
Jazz Instrument
Romancing 123.452
Pop Vocal Romancing 49.877
Page 10
Artist and Usage (1)
Dependency analysis Binomial exact test on usages and artists with 10 reviews
Artist Usage p value
AFI Waking Up 0.03252
Black Sabbath
At Work 0.00028
Celine Dion Romancing 0.02499
Dream Theater
Listening 0.01862
Metallica Waking Up 0.03252
Nirvana_(USA)
Going to Sleep
0.01862
Page 11
Artist and Usage (2)
Usage ProfilesUsage distributions of 10 most-reviewed artists
Page 12
Clustering on co-occurrences
Some usages appear to be related e.g. “Exercising” and “Cleaning the
House”
Q3: Can the usages form meaningful superclasses base on their co-occurrences with genre, artist and album titles?
Page 13
Clusters from genre-usage co-occurrences
RomancingGetting ready to go
outExercisingWaking up
Hanging out with friends
At workDriving
ListeningGoing to sleep
Reading or studyingCleaning the house
Relaxing
Stimulating
Page 14
Clusters from artist-usage co-occurrences
Going to sleepListening
Reading or studyingRomancing
At workExercisingWaking up
Getting ready to go out
DrivingCleaning the house
Hanging out with friends
Relaxing
Stimulating
Page 15
Clusters from album-usage co-occurrences
Going to sleepReading or studying
ListeningExercisingWaking up
At workCleaning the house
DrivingHanging out with
friendsGetting ready to go
outRomancing
Relaxing
Stimulating
Page 16
To summarize …
Stimulating:Exercising, Waking up, At work, Driving,Hanging out with friends, Getting ready to go out
Relaxing: Going to sleep, Reading or studying
Discrepant: Relaxing
Stimulating
separate
Listening 2 1 0
Romancing 1 1 1
Cleaning the house
0 2 1
Page 17
Data Limitations
Only from one websiteUsage choices are predefinedSome usages are ambiguousInterpretations vary across usersOnly one usage per review
can’t see how individual users group usages
Page 18
Conclusions
Usage: another facet of music similarityComplementary to artist and genre similarity
Consistent superclasses of usagesMeaningful user-generated metadataNew task / query for MIREX
Further investigation is warrantedLarger scale dataset from multiple websitesConnect to audio features
Page 19
Questions?
IMIRSEL
Thank you!
THE ANDREW W. MELLON FOUNDATION
Page 20
Usage Categories and Counts
Usage Count
Usage Count
Driving 1,349 Waking up 271
Hanging With Friends 1,215 Going to Sleep 269
Listening 592 Cleaning the House
230
Romancing 492 At Work 188
Reading or Studying 447 With Family 35
Getting ready to go out
378 Sleeping 15
Exercising 291 TOTAL 5,772
Page 21
Agenda
MotivationUsage metadata in www.epinions.comResearch QuestionsAnalysis
Genre and usageArtist and usageClustering on co-occurrences
Conclusions
Page 22
Clustering on co-occurrences
Some usages appear to be related e.g. “Exercising” and “Cleaning the House”
Q3: Can the usages form meaningful superclasses base on their co-occurrences with genre, artist and album titles?
Facets with multiple usages
number reviews
Genres 12 1,980Artists 368 1,451
Album titles 366 974