Slides of my PhD thesis defense presentation (16-12-2022, Barcelona, Spain). I obtained the obtained the grade of Excellent and I have been awarded with the Mention "Cum Laude".
Labelling Requirements and Label Claims for Dietary Supplements and Recommend...
Lorenzo Porcaro PhD Defense
1. Assessing the Impact of Music Recommendation
Diversity on Listeners
Lorenzo Porcaro
Supervisors
Dr. Emilia Gómez and Dr. Carlos Castillo
Committee
Dr. Christine Bauer (Utrecht University)
Dr. Perfecto Herrera (Universitat Pompeu Fabra)
Dr. Mounia Lalmas-Roelleke (Spotify)
----------------------------------------------------------------------
Dr. Dmitry Bogdanov (Universitat Pompeu Fabra)
Dr. Sergio Oramas (Pandora) 16/12/2022
1
2. Molino, J., & Ayrey, C. (1990). Musical Fact and the Semiology of Music. Music Analysis, 9(2), 105–111; 112–156.
Semiology (“study of signs”) →Discipline that studies the phenomena of signification and communication.
2
5. Goldberg, D., Nichols, D., Oki, B. M., & Terry, D. (1992).
Using collaborative filtering to Weave an Information
tapestry. Communications of the ACM, 35(12),
61–70. https://doi.org/10.1145/138859.138867 5
10. Thesis Statement
We provide empirical evidence of the function that diversity
plays in mediating the relationships between music
recommendations and listeners. Connecting the
measurement, perception and impact of diversity we
deepen the understanding of recommender systems' role in
shaping today's music listening experience.
10
13. Music Recommendation Diversity
Measurements
Porcaro, L., & Gomez, E. (2019). 20 Years of Playlists: a Statistical Analysis on Popularity and Diversity.
Proceedings of the 20th International Symposium on Music Information Retrieval, ISMIR 2019, July, 4–11.
Porcaro, L., Castillo, C., & Gómez, E. (2021). Diversity by Design in Music Recommender Systems.
Transactions of the International Society for Music Information Retrieval, 4(1), 114–126.
13
14. 14
Music Recommender System Diversity
Item Diversity User Diversity
Behavioural
Diversity
Item Features
User
Characteristics
Content Source
Perceived Diversity
Poietic Domain Esthetic Domain
- Audio Signal
- Metadata
- Taxonomies
- ...
- Demographics
- Personality Traits
- Personal Values
- ...
Exposure Exposure
16. 16
1998 C
C
C
C
C
C
2010 2011 2012 2013 2015 2018
/
AOTM¹
# tracks: 972K
# playlist: 100K
type: user-generated
catalogue: user
CORN²
# tracks: 15K
# playlist: 75K
type: radio playlist
catalogue: radio
SPOT³
# tracks: 2M
# playlist: 175K
type: user-generated
catalogue: streaming
DEEZ⁴
# tracks: 227K
# playlist: 82K
type: user-generated
catalogue: streaming
¹ McFee, B., & Lanckriet, G. “Hypergraph models of playlist dialects”. Proceedings of the 13th International Society for Music Information Retrieval Conference 343–348. 2012
² S. Chen, J.L. Moore, D. Turnbull, and T. Joachims. “Playlist prediction via metric embedding”, Proc. of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ’12, 2012
³ M. Pichl, E. Zangerle, and G.Specht. “Towards a Context-Aware Music Recommendation Approach: What is Hidden in the Playlist Name?”, Proc. of the 15th IEEE International Conference on Data Mining Workshop
⁴ Crawled in-house
21. 21
Popularity Tags
Average frequency of its tracks Average distance between its tracks’ tags
Shannon / Simpson / Gini index Descriptive / Gini / Qualitative
Playlist Diversity
Playlist Dataset Diversity
23. 23
Radio playlists VS user-generated playlist
Radio more balanced in terms of popularity
and tags.
24. 24
Radio playlists VS user-generated playlist
Radio more balanced in terms of popularity
and tags.
Oldest playlist (pre-streaming) VS newest
playlist (streaming)
25. Radio playlists VS user-generated playlist
Radio more balanced in terms of popularity
and tags.
Oldest playlist (pre-streaming) VS newest
playlist (streaming)
Oldest lower popularity (especially user
generated), and more balanced in terms of tags.
25
27. ❏ Diversity metrics as tools for comparative analysis.
❏ Playlist creation strategy may reflect technological contexts.
27
28. ❏ Diversity metrics as tools for comparative analysis.
❏ Playlist creation strategy may reflect technological contexts.
❏ There is no average user in the real-world.
28
29. ❏ Diversity metrics as tools for comparative analysis.
❏ Playlist creation strategy may reflect technological contexts.
❏ There is no average user in the real-world.
❏ The concept of diversity may vary among different people.
29
30. Music Recommendation Diversity
Perceptions
Porcaro, L., Gómez, E., & Castillo, C. (2022). Perceptions of Diversity in Electronic Music: The Impact of
Listener, Artist, and Track Characteristics. Proc. ACM Hum.-Comput. Interact., 6(CSCW1).
Porcaro, L., Gómez, E., & Castillo, C. (2022). Diversity in the Music Listening Experience : Insights from Focus
Group Interviews. Proceedings of the 2022 ACM SIGIR Conference on Human Information Interaction and
Retrieval (CHIIR ’22), 272–276.
30
39. Track Diversity
Input: hand-crafted MIR features
Input type: numerical
Metric: cosine distance
Artist Diversity
Input: gender, birthplace, skin tone, debut year
Input type: categorical
Metric: Goodall distance
40. 2 2
2
2
2
2
List B
Average: div(List A) > div(List B)
→ all distances are equally important (utilitarian)
avg(di
) = 2
min(di
) = 2
3
4
3
3
4 1
List A
avg(di
) = 3
min(di
) = 1
41. 3
4
3
3
4 1
2 2
2
2
2
2
List A
List B
Average: div(List A) > div(List B)
→ all distances are equally important (utilitarian)
Minimum: div(List B) > div(List A)
→ focus on closest distance (egalitarian)
avg(di
) = 3
min(di
) = 1
avg(di
) = 2
min(di
) = 2
43. Participants (N=115)
61% - Aged between 18 and 35
73% - Male
73% - Europe or North America
85% - Bachelor’s degree or higher
60% - Skin tone type I-II (white skin)
44. Participants (N=115)
61% - Aged between 18 and 35
73% - Male
73% - Europe or North America
85% - Bachelor’s degree or higher
60% - Skin tone type I-II (white skin)
WEIRD (Western, Educated,
Industrialized, Rich, and
Democratic) societies.
45. Participants (self-declared) Musical Background
❖ Playing, DJ-ing, or producing
❖ Varied listening habits
❖ Mostly electronic music listeners
❖ Within electronic music, listen to different styles
46. RQ1. To what extent
tracks’ audio-features
and artists’ attributes
can be used to assess
the perceived
diversity?
RQ2. To what extent
domain knowledge
and familiarity
influence participants’
perceptions of
diversity?
47. Can audio-based features be used to determine diversity?
Sometimes: when ≥ 85% participants agree on the most
diverse list ...
… the model identifies the same list as more diverse.
48. Can artists’ attributes be used to model a degree of diversity?
Answers of participants aged <35, and WEIRD participants
→ moderate-to-substantial agreement* with Goodall.
* ≥ .55 Light’s Kappa (average pairwise Cohen’s Kappa)
49. RQ1. To what extent
tracks’ audio-features
and artists’ attributes
can be used to assess
the perceived
diversity?
RQ2. To what extent
domain knowledge
and familiarity
influence participants’
perceptions of
diversity?
50. RQ2. To what extent domain knowledge and familiarity influence
participants’ perceptions of diversity?
Listeners with
low domain knowledge
51. RQ2. To what extent domain knowledge and familiarity influence
participants’ perceptions of diversity?
Listeners with
low domain knowledge
Listeners with
high domain knowledge
52. ❏ Diversity-aware music recommendations are good for users with
less domain knowledge, but less useful for more specialized users.
❏ New insights for designing socially-relevant diversity-aware music
recommendations.
53. ❏ Diversity-aware music recommendations are good for users with
less domain knowledge, but less useful for more specialized users.
❏ New insights for designing socially-relevant diversity-aware music
recommendations.
54.
55. Interviews
❖ 14 participants
❖ 7 semi-structured interview (2-3 participants)
❖ Duration ~ 30’ - 60’
❖ Structure:
1. Music list diversity assessments
2. Experience with music recommender systems
57. Less knowledge (newcomers)
→ diversity assessment harder
→ rely more generic features
“People who don’t know much about a
particular genre, probably would agree on
some things that are a bit more generic.”
59. More knowledge (experts)
→ easier to categorize
→ more receptive to details BUT more bias
“I can feel like I can make a better decision of
what is diverse [...] but then there is kind of a
bias that comes based on the fact that I like
this music a lot.”
60. Diversity as a tool for deconstructing stereotyped views of a music culture.
61. Diversity as a tool for deconstructing stereotyped views of a music culture.
“The electronic music that I went to listen to and
that I liked before the survey was predominantly
white male, which I suppose is still what is
predominant in the industry to some extent.
62. Diversity as a tool for deconstructing stereotyped views of a music culture.
“I realized while making the questionnaire that I
had a very strict let’s say definition about
electronic music myself.”
64. Algorithmic recommendation diversity help listeners in discovering new
facets of disliked music genres.
“I never liked EDM but [...] algorithms presented
to me different tracks, and I found myself
listening to it, and noticing the differences
within this genre. In the end, I started listening
to it more often.”
65. Music Recommendation Diversity
Impacts
Porcaro, L., Gómez, E., and Castillo, C. (under review). Assessing the Impact of Music Recommendation
Diversity on Listeners: A Longitudinal Study. ACM Transactions of Recommender Systems (TORS).
65
67. Prescreening
First step (Prolific)
❖ Age: 18-42
❖ Nationality: Italy, Spain, Portugal
❖ Gender and sex: No restrictions
❖ Highest education level: No restrictions
❖ Number of Prolific previous submissions: > 20
❖ Prolific approval rate: 90%
68. Prescreening (cont.)
Second step (PsyToolkit)
❖ Taste variety: No restrictions
❖ Electronic Music listening frequency: Not very often
❖ Average daily listening time: > 1 hour
74. PRE
LS
Week 5 Week 6 Week 7 Week 8
Prescreening
EMF
Weeks 1-4
EMF EMF EMF
COND
EMF
LS
LS
LS
LS
LS
LS
LS
EMF = Electronic Music Feedback Questionnaire
LS = Listening Session
Week 5 Week 6 Week 7 Week 8
Weeks 1-4 Weeks 9-12
76. PRE
LS
Week 5 Week 6 Week 7 Week 8
Prescreening
EMF
Weeks 1-4 Weeks 9-12
EMF EMF EMF EMF
COND POST
EMF
LS
LS
LS
LS
LS
LS
LS
EMF = Electronic Music Feedback Questionnaire
LS = Listening Session
83. PRE
LS
Week 5 Week 6 Week 7 Week 8
Prescreening
EMF
Weeks 1-4 Weeks 9-12
EMF EMF EMF EMF
COND POST
EMF
LS
LS
LS
LS
LS
LS
LS
EMF = Electronic Music Feedback Questionnaire
LS = Listening Session
84.
85.
86. 1. High Diversity group engaged more with the playlists than
Low Diversity group.
2. Low Diversity group liked more the music listened to than
the High Diversity group.
3. High Diversity and Low Diversity had the same level of
familiarity with the music listened to.
86
87. 1. High Diversity group engaged more with the playlists than
Low Diversity group.
2. Low Diversity group liked more the music listened to than
the High Diversity group.
3. High Diversity and Low Diversity had the same level of
familiarity with the music listened to.
87
88. 1. High Diversity group engaged more with the playlists than
Low Diversity group.
2. Low Diversity group liked more the music listened to than
the High Diversity group.
3. High Diversity and Low Diversity had the same level of
familiarity with the music listened to.
88
89. PRE
LS
Week 5 Week 6 Week 7 Week 8
Prescreening
EMF
Weeks 1-4 Weeks 9-12
EMF EMF EMF EMF
COND POST
EMF
LS
LS
LS
LS
LS
LS
LS
EMF = Electronic Music Feedback Questionnaire
LS = Listening Session
90.
91.
92. ❖ The implicit association with Electronic Music tends
towards less extreme valence.
❖ The openness in listening to Electronic Music increases.
92
93. ❖ The implicit association with Electronic Music tends
towards less extreme valence.
❖ The openness in listening to Electronic Music increases.
* Slight difference between PRE-COND and PRE-POST
** No particular influence by the degree of exposure diversity
93
94. RQ1. To what extent
listeners’ implicit and
explicit attitudes
towards an unfamiliar
music genre can be
affected by exposure to
music recommendations?
RQ2. What is the
relationship between
music recommendation
diversity and the
impact on listeners’
attitudes?
95. RQ2. What is the
relationship between
music recommendation
diversity and the
impact on listeners’
attitudes?
RQ1. To what extent
listeners’ implicit and
explicit attitudes
towards an unfamiliar
music genre can be
affected by exposure to
music recommendations?
100. Findings
Music recommendation diversity
➔ Measurements
◆ Differences in playlist creation strategies
➔ Perceptions
◆ Agreement between computational metrics and perceived diversity
➔ Impacts
◆ Impacts of music recommendation diversity on listeners’ attitudes
100
104. Contributions & Awards
6 papers (2 journal, 2 conference, 2 workshop)
3 EU-funded Project (divinAI, TROMPA, MusicalAI) + 7 papers
Software and datasets public available
Women in RecSys Journal Paper of the Year (RecSys 2022)
SIGIR Student Travel Award (CHIIR 2022)
The Sónar+D Innovation Challenge 2020
104
105. Dissemination
European Researchers’ Night
Oracle4Girls initiative
MUTEK Digital Art and Music Festival Symposium 2021 & 2022
ORION Open Science Podcast
Interview at TV3 (Catalan Television Broadcaster)
Article featured in the UPF Institutional Website
HUMAINT project, Joint Research Centre, European Commission
105
106. Assessing the Impact of Music Recommendation
Diversity on Listeners
Lorenzo Porcaro
Supervisors
Dr. Emilia Gómez and Dr. Carlos Castillo
Committee
Dr. Christine Bauer (Utrecht University)
Dr. Perfecto Herrera (Universitat Pompeu Fabra)
Dr. Mounia Lalmas-Roelleke (Spotify)
----------------------------------------------------------------------
Dr. Dmitry Bogdanov (Universitat Pompeu Fabra)
Dr. Sergio Oramas (Pandora) 16/12/2022
106