Leveraging Social Media Sources to Generate Personalized Music Playlists

EC-Web 2012 - 13th International Conference on Electronic Commerce and Web Technologies
Vienna (Austria), 04.09.2012

Cataldo Musto, Giovanni Semeraro, Pasquale Lops, Marco de Gemmis, Fedelucio Narducci

Leveraging Social Media Sources to
Generate Personalized Music Playlists

Some stats
28,000,000 songs available on iTunes Store (*)
around 31,000 hours of music

a typical user spends 1.5 hours for day listening to music
=
56 years
to listen to the whole iTunes Library
(*) http://www.digitalmusicnews.com/permalink/2012/120425itunes

C. Musto, G. Semeraro, P. Lops, M. de Gemmis, F. Narducci.
Leveraging Social Media Sources to Generate Personalized Music Playlists - EC-WEB 2012 - 04.09.12

exponential growth
of the available music

online music libraries totally
changed our listening habits

Information Overload


paradox of choice
(Barry Schwartz, TED talk “Why more is less”)


what music should I listen to?

solution

personalization.

solution

personalized music playlists

Is this something new?
No.


Amazon.com

Recommendations


Genius @iTunes

Recommendations


All the state of the art
platforms share an
important drawback.


need for
explicit
information
about
user interests.


training is a bottleneck.

social media

changed the rules for information
management and knowledge acquisition

social media

give new opportunities for modeling and
gathering information about user preferences

Facebook

user preferences in music.

Our contribution
Play.me


Play.me
personalized music playlists
• Goal
• To provide users with personalized music playlists
• Insights
• Extraction of explicit user preferences from Facebook
• Playlist creation by enriching explicit user preferences.
• New artists are added to those explicitly extracted from
Facebook

• Comparison of two enrichment techniques

Play.me
architecture


Myusic
pre-processing

• Crawling from Last.fm
• Public API
• Content-based features
• Name of the artist + Social tags
• Noise processing
• Information locally stored

Myusic
pre-processing

Sigur Ròs tag cloud from Last.fm


Myusic
data extraction from Facebook


Myusic

explicit preferences

Myusic

implicit preferences

Play.fm
enrichment

• Rationale
• Given a set of explicit preferences extracted from
Facebook
• Play.me enrichs this set
• Extraction of artists similar to those the user
explicity likes


Play.fm
enrichment example

Coldplay extracted from Facebook
enrichment

radiohead red hot chili peppers kings of leon


Play.me
playlist

Most popular songs of the artists extracted from Last.fm (as well as
those added through the enrichment) are proposed to the user.

let’s go
deeper

Play.fm
enrichment

• Comparison of two approaches
•
Content-based strategy
• Distributional Models

• Linked Data


Play.fm
enrichment based on Distributional Models

• Content-based strategy
• Each artist is represented through a set of
tags
• Each artist is represented as a point in a
semantic geometrical space
• Distributional Models
• Similarity calculations to extract the most
similar artists.

distributional models
insight
by analyzing large corpus of textual data it is possible
to infer information about the usage (about the meaning)
of the terms.



“meaning
is its use”
L.Wittgenstein
(Austrian philosopher)


insight
by analyzing large corpus of textual data it is possible
to infer information about the usage (about the meaning)
of the terms.
example


term/context matrix (WordSpace)

c1 c2 c3 c4 c5 c6 c7 c8 c9

t1 ✔ ✔ ✔ ✔
t2 ✔ ✔ ✔ ✔

t3 ✔ ✔ ✔
t4 ✔ ✔ ✔ ✔

Cataldo Musto - Enhanced Vector Space Models for Content-based Recommender Systems - Ph.D. defense - University of Bari Aldo Moro, Italy - 08.06.12

beer vs. glass: good overlap

c1 c2 c3 c4 c5 c6 c7 c8 c9

t1 ✔ ✔ ✔ ✔
t2 ✔ ✔ ✔ ✔

t3 ✔ ✔ ✔
t4 ✔ ✔ ✔ ✔


beer vs. spoon: no overlap

c1 c2 c3 c4 c5 c6 c7 c8 c9

t1 ✔ ✔ ✔ ✔

t2 ✔ ✔ ✔ ✔

t3 ✔ ✔ ✔

t4 ✔ ✔ ✔ ✔


rock vs. post rock = good overlap

a1 a2 a3 a4 a5 a6

rock ✔ ✔ ✔
post rock ✔ ✔
jazz ✔

classical ✔ ✔ ✔

rock vs. classical = no overlap

a1 a2 a3 a4 a5 a6

rock ✔ ✔ ✔

post rock ✔ ✔

jazz ✔

classical ✔ ✔ ✔

WordSpace
example

rock
post-rock

jazz

classical

representation of documents (*)
can be inferred by combining the representation of
the terms (**) occurring in the document.

(*) documents = artists
(**) terms = tags


term/context matrix (DocSpace)

c1 c2 c3 c4 c5 c6 c7 c8 c9

t2 ✔ ✔ ✔ ✔

t3 ✔ ✔ ✔

d1 ✔ ✔ ✔ ✔ ✔

Play.fm

Coldplay
Radiohead
Kings of Leon

Lady Gaga

Play.fm

input: vector space representation

output: artists with the highest cosine similarity

radiohead the killers kings of leon

Linked Open Data Cloud


Linked Open Data Cloud

Structured
(RDF)
representation
of the information
stored in Wikipedia.

Play.fm
enrichment based on Linked Data

Coldplay play Alternative Rock

Play.me
RDF triple

Relationships are explictly encoded in RDF.


Play.fm

• Linked Open Data Cloud
• Each artist is mapped on a DBpedia node.
• univocal URI
• Relationship between artists (nodes) are explicitly
encoded
• e.g. genre, artist category, etc.
• Use of SPARQL to extract artists (nodes) that
share the same features

Play.fm

input: SPARQL query

output: artists sharing the same properties

radiohead the smiths the verve

recap
enrichment process
input: artist output: similar artists
coldplay the smiths

Linked Data
radiohead

the verve

kings of leon
Distributional Models

radiohead


experimental
evaluation.


experimental design

• Experiment
• Which one is the enrichment technique that
can provide users with the best playlists ?


experimental design
settings

• 30 users
• Heterogeneous musical knowledge
• Last.fm crawl: 228,878 artists
• Extraction & Recommendation step
• 325 artists extracted
• 11 per user, on average

experimental setup

Given a playlist, each user can freely express her own
feedback (like/dislike) on the proposed tracks.

experimental setup

Experiment repeated three times (one run with Linked Data enrichment,
another one with Distributional Models, one with a simple baseline based on
popularity).

experimental setup

Users were unaware of the adopted conﬁguration.


experimental design
results
76.3
80

75.2 Linked Data
Baseline (Popularity)
73.75
69.7
67.5

65.9
64.6
61.25 63.2

58 58 58
55
n=1 n=2 n=3
n = number of artists added for each extracted artist

experimental design
results
76.3
80

75.2 Linked Data
73.75
69.7
67.5

65.9
64.6
61.25 63.2

58 58 58
55
n=1 n=2 n=3
distributional models overcome linked data

experimental design
results
76.3
80

75.2 Linked Data
73.75
69.7
67.5

65.9
64.6
61.25 63.2

58 58 58
55
n=1 n=2 n=3
precision in distributional models drops down more rapidly

experimental design
results
76.3
80

75.2 Linked Data
73.75
69.7
67.5

65.9
64.6
61.25 63.2

58 58 58
55
n=1 n=2 n=3
good results for baseline, as well (poor music knowledge?)

Play.me
recap of the experiment

• Play.me
• Platform for generating personalized music playlists
•Extraction of favourite artists from Facebook

• Experiments
• Enrichment based on distributional models
overcomes linked data-based one in a signiﬁcant
way

conclusions.


both enrichment techniques
overcome the baseline

overcome linked data

future research.

merging different
enrichment techniques

evaluation with user-based metrics
(serendipity, novelty, unexpectedness)

modeling context.

questions?

Leveraging Social Media Sources to Generate Personalized Music Playlists

Recommended

Recommended

More Related Content

Viewers also liked

Viewers also liked (17)

More from Cataldo Musto

More from Cataldo Musto (20)

Recently uploaded

Recently uploaded (20)

Leveraging Social Media Sources to Generate Personalized Music Playlists

Editor's Notes