Finding a Path Through the Juke Box: The Playlist Tutorial

Finding a path through the
Juke Box
The Playlist Tutorial

Ben Fields, Paul Lamere
ISMIR 2010

“I still maintain that music is
the best way of getting the
self-expression job done.”
Nick Hornby

Overview
• Introduction
• Brief History of playlists
• Aspects of a good playlist
• Automatic generation of playlists
• Survey of automatic playlisters
• Evaluating playlists
• An evaluation of various playlisting services
• The future of playlisting

3

Goals

• Understand where and why playlists are
important
• Understand current and past methods of
playlist construction
• Understand the whys and hows of various
evaluation methods

4

What is a playlist?
• mixtape
• prerecorded DJ set/mix CD
• live DJ set (typically mixed)
• radioshow logs
• an album
• functional music (eg. Muzak)
• any ordered list of songs?
6

What is a playlist?

we deﬁne a playlist as a set of songs meant
to be listened to as a group, usually with an
explicit order

7

Why is playlisting
important?
• Ultimately, music is consumed through
listening
• An awareness of this act of listening is critical
to successful MIR application
• The playlist is a formalization of this listening
process
• Playlists have a traditional revenue model for
artists and labels (e.g. radio)

8

Brief History of
Playlists

Mixed Concert Programs

• Marks the beginnings international
combinations of music from multiple
composers
• Begins circa 1850 in London
• The idea of a set of music being curated
begins to form

From miscellany to homogeneity in concert programming
William Weber 10

Early Broadcast Media
• moving the ethos of the earlier period onto the
radio

• biggest changes are technology

• broadcast = larger simultaneous audience

• phonograph brings recorded music

• initial broadcasts (eg. 1906 - Fessenden) as
publicity stunts

• ﬁrst continuous broadcast 1920 - Frank Conrad

The slow pace of rapid technological change: Gradualism and punctuation in technological change
Daniel A. Levinthal 11

Rock On the Radio
• radio as a medium begins to push certain
genres, especially rock and roll and r ‘n’ b
• playlist ﬁrst used to describe (unordered)
sets of songs
• personality driven
• John Peel
• Casey Kasem
Last Night A DJ Saved My Life; The history of the disc jockey Finding an alternative: Music programming in US college radio
Bill Brewster and Frank Broughton Tim Wall 12

Disco & Hip-Hop
emergence of the club DJ
• DJ as Disco nightclubs, with a mixer and two turntables,
saw the birth of the idea of continuous mixing

• DJs wanted dancers to not notice song transitions, and
techniques such as beat matching and phrase
alignment were pioneered

• Hip-Hop saw this idea pushed further, as DJs became live
remixers, turning the turntable into an instrument

• At the same time, club DJs started to become the top
billing over live acts, the curator becoming more of a draw
than the artist

Last Night A DJ Saved My Life; The history of the disc jockey
Bill Brewster and Frank Broughton 13

The Playlist Goes Personal
• The emergence of portable audio devices drives the
popularity of cassette tapes

• This in turn leads to reordering and combining of disparate
material into mixtapes

• Mixtapes themselves are traded and distributed socially,
providing a means for recommendation and discovery

• In hip-hop, mixtapes served as the ﬁrst recordings of new
DJs featuring novel mixes and leading to current
phenomenon of Mix [CD|set|tape] (now on CD or other
digital media)

Investigating the Culture of Mobile Listening: From Walkman to iPod
Michael Bull 14

Now With Internet
• The Web’s increase in popularity and MP3
audio compression allow for practical sharing
of music of the Internet
• This brings the mixtape for physical sharing to
non-place sharing.
• Streaming-over-internet radio emerges
• Playlists on the cloud: play.me, spotify, etc.

Remediating radio: Audio streaming, music recommendation and the discourse of radioness
Ariana Moscote Freire 15

Aspects of a good
playlist

Aspects of a good Playlist
To me, making a tape is like writing a letter — there's a lot
of erasing and rethinking and starting again. A good
compilation tape, like breaking up, is hard to do. You've got
to kick off with a corker, to hold the attention (...), and then
you've got to up it a notch, or cool it a notch, and you can't
have white music and black music together, unless the
white music sounds like black music, and you can't have
two tracks by the same artist side by side, unless you've
done the whole thing in pairs and...oh, there are loads of
rules. - Nick Hornby, High Fidelity

17

Factors affecting a good playlist

• The songs in the playlist - including the listener’s
familiarity with and preference for the songs

• The level of variety and coherence in a playlist

• The order of the songs:

• The song transitions

• Overall playlist structure.

• Other factors: serendipity, freshness,
‘coolness’,

• The Context
Learning Preferences for Music Playlists
A.M. de Mooij and W.F.J. Verhaegh 18

Factors affecting a good playlist

Survey with 14 participants

Learning Preferences for Music Playlists
A.M. de Mooij and W.F.J. Verhaegh
19

Factors affecting preference

• Musical taste - long term slowly evolving commitment to a genre

• Recent listening history

• Mood or state of mind

• The context:
listening, driving, studying,
working, exercising, etc.

• The Familiarity

• People sometimes prefer to listen to the familiar songs that they
like less than non-familiar songs
• Familiarity signiﬁcantly predicts choice when controlling for the
effects of liking, regret, and ‘coolness’

I Want It Even Though I Do Not Like It: Preference for Familiar but Less Liked Music Learning Preferences for Music Playlists
Morgan K. Ward, Joseph K. Goodman, Julie R. Irwin A.M. de Mooij and W.F.J. Verhaegh
20

Coherence
Organizing principals for mix help requests

• Artist / Genre / Style
• Song Similarity
• Event or activity
• Romance
• Message or story
• Mood
• Challenge or puzzle
• Orchestration
• Characteristic of the mix recipient
• Cultural References

‘More of an Art than a Science’: Supporting the Creation of Playlists and Mixes
Sally Jo Cunningham, David Bainbridge, Annette Falconer 21

Coherence

• Artist / Genre / Style “acoustic-country-folk type stuff”,

• Song Similarity
• Event or activity
• Romance
• Mood
• Orchestration


Coherence


• Song Similarity
• Event or activity “anti-Valentine mix”

• Romance
• Mood
• Orchestration


Coherence


• Song Similarity

• Romance
a mix with the title “‘quit being a
• Message or story douche’, ’cause I’m in love with you.

• Mood
• Orchestration


Coherence


• Song Similarity

• Romance

• Mood
song whose title is a question?
• Orchestration


Coherence


• Song Similarity

• Romance

• Mood
• Orchestration songs where the singer hums for a little bit



Coherence


• Song Similarity

• Romance

• Mood
• Orchestration songs where the singer hums for a little bit

• Cultural References “songs about superheroes


“People have gotten used to
listening to songs in the order
they want, and they'll want to
continue to do so even if they
can't get the individual songs
from ﬁle-trading programs.”
Phil Leigh

Ordering Principals
• Bucket of similars, genre

• Acoustic attributes such as tempo, loudness, danceability

• Social attributes such popularity, ‘hotness’

• Mood attributes (‘sad’ to ‘happy’)

• Theme / Lyrics

• Alphabetical

• Chronological

• Random

• Song transitions

• Novelty orderings

23

Novelty ordering

0 We Wish You A Merry Christmas - Weezer
1 Stranger Things Have Happened - Foo Fighters
2 Dude We're Finally Landing - Rivers Cuomo
3 Gotta Be Somebody's Blues - Jimmy Eat World
4 Someday You Will Be Loved - Death Cab For Cutie
5 Dancing In The Moonlight - The Smashing Pumpkins
6 Take The Long Way Round - Teenage Fanclub
7 Don't Make Me Prove It - Veruca Salt
8 The Sacred And Profane - Smashing Pumpkins, The
9 Everything Is Alright - Motion City Soundtrack
10 Trains, brains & rain - The Flaming Lips
11 No One Needs To Know - Ozma
12 What Is Your Secret - Nada Surf
13 The Spark That Bled - Flaming Lips, The
14 Defending The Faith - Nerf Herder

24

Where song order rules
The Dance DJ
• For the Dance DJ - song order and transitions are especially
important

• Primary goal: make people dance

• How?

• Selecting

• tracks that mix well

• takes the audience on a journey

• audience feedback is important

• Mixing

• seamless song transitions
Is the DJ an Artist?
Is a mixset a piece of art?
Hang the DJ: Automatic Sequencing and Seamless Mixing of Dance-Music Tracks
Dave Cliff Publishing Systems and Systems Laboratory HP Laboratories Bristol HPL-2000-104 9th August, By BRENT SILBY 25
2000*

Tempo Trajectories

Warmup Cool down Nightclub

hpDJ: An automated DJ with floorshow feedback
Dave Cliff Digital Media Systems Laboratory HP Laboratories Bristol 26

Coherence
Song to Song

Beat Matching and Cross-fading
hpDJ: An automated DJ with floorshow feedback
Dave Cliff Digital Media Systems Laboratory HP Laboratories Bristol 27

Don’t underestimate the power of the shufﬂe

THE SERENDIPITY SHUFFLE
Tuck W Leong, Frank Vetere , Steve Howard 28


laugh-out-loud pleasurable



white-knuckle ride


“...teaches me connections between
disparate kinds of music and the inﬁnite
void. I understand the universe better”



...forge(ing) new connections
between my heart and my ears



each randomly-sequenced track like an aural postcard



had made me re-examine things I thought I
knew about my favourite music



...hear(ing) songs that I haven’ t heard for years and
fall(ing) in love with them again



Random shufﬂe can turn large music libraries
into an ‘Aladdin’ s cave’ of aural surprises


...the random effect delivers a sequence of music so perfectly
thematically 'in tune'that (it) is quite unsettling


Serendipity of the shufﬂe

Finding meaningful experience in chance encounters

• Serendipity can improve the listening experience

• Choosing songs randomly from a personal
collection can yield serendipitous listening

• Drawing from too large, or too small of a
collection reduces serendipity


People like shufﬂe play

People shufﬂe genres, albums and playlists
Randomness as a resource for design

Playlist tradeoffs

Variety Coherence

Freshness Familiarity

Surprise Order

Different listeners have different optimal settings
Mood and context can affect optimal settings

31

Playlist Variety
A good playlist is not a bag of similar tracks

32

Playlisting is not Recommendation

Recommendation Playlist

Primarily for music discovery Primarily for music listening

Minimize familiar artists Familiar artists in abundance

Order not important Order can be critical

Rich contexts - party,
Limited Context (shopping)
jogging, working, gifts

However, playlists may be better vector for music
discovery than traditional recommendation
33

Playlisting nuts and
bolts
formats and rules

34

Playlist formats

• Lots of formats - Some notable examples:

• M3U - simple list of ﬁles - one per line

• XSPF - ‘spiff’ - XML based format

• The Playback Ontology

• Resources:

• http://microformats.org/wiki/audio-info-formats

• http://lizzy.sourceforge.net/docs/formats.html

• http://gonze.com/playlists/playlist-format-survey.html

35

Example XSPF
<?xml version="1.0" encoding="UTF-8"?>
<playlist version="1" xmlns="http://xspf.org/ns/0/">
<trackList>
<track>
<location>http://example.com/song_1.mp3</location>
<creator>Led Zeppelin</creator>
<album>Houses of the Holy</album>
<title>No Quarter</title>
<annotation>I love this song</annotation>
<duration>271066</duration>
<image>http://images.amazon.com/images/P/B000002J0B.jpg</image>
<info>http://example.com</info>
</track>
<track>
<location>http://example.com/song_1.mp3</location>
<creator>Led Zeppelin</creator>
<album>ii</album>
<title>No Quarter</title>
<annotation>This one too</annotation>
<duration>271066</duration>
<image>http://images.amazon.com/images/P/B000002J0B.jpg</image>
<info>http://example.com</info>
</track>
</trackList>
</playlist>

36

The Playback Ontology
The Play Back Ontology provides basic concepts and properties for describing
concepts that are related to the play back domain, e.g. a playlist,play back and
skip counter, on/ for the Semantic Web.

http://smiy.sourceforge.net/pbo/spec/playbackontology.html http://smiy.wordpress.com/2010/07/27/the-play-back-ontology/

Modeling items in the playlist by extending the ordered list ontology


Expressing similarity and creation provenance


Survey of playlisting
systems and tools

Social

Manual Automated

Non-Social 41

Social

Manual Automated

Non-Social 42

Manual Non-Social

43

Rush: Repeated Recommendations on Mobile
Devices

Rush: Repeated Recommendations on Mobile Devices
Dominikus Baur, Sebastian Boring, Andreas Butz 44

Playlist creation tools

45

Do people use Smart Playlists?

30

22.5
Percent

15

7.5

0
No iTunes Never 1 to 5 6 to 10 11 to 20 21 to 100 over 100

Informal poll with 162 respondents
46

Social

Manual Automated

Non-Social 47

Automated Non-Social

48


49

MOG

49

Mood Agent

• Use sliders to set levels
of 5 ‘moods’:

• Sensual

• Tender

• Happy

• Angry

• Tempo

50

Visual Playlist Generation on the Artist Map

Visual Playlist Generation on the Artist Map
Van Gulick, Vignoli
52

GeoMuzik

GeoMuzik: A geographic interface for large music
collections: Òscar Celma, Marcelo Nunes 54

Using visualizations to build playlists

MusicBox: Mapping and visualizing music collections
Anita Lillie’s Masters Thesis at the MIT Media Lab

55

Search Inside the Music

Using 3D Visualizations to explore and discover music.
Paul Lamere and Doug Eck 56

Social

Manual Automated

Non-Social 57

Automated Social

58

Automated Social

Last.fm
58

Automated Social

59

DMCA Radio
US rules for Internet streaming radio
• In a single 3 hour period:

• No more than three songs from the same
recording

• No more than two songs in a row, from the
same recording

• No more than four songs from the same artist or
anthology

• No more than three songs in a row from the
same artist or anthology
Note that there are no explicit rules that limit skipping

60

Terrestrial Radio Programming

61

Radio station programming rules

• Divide the day into a set of 5 (typically) ‘dayparts’.:
Mid-6A, 6A-10A, 10A-3P, 3P-7P, and 7P-12Mid

• For each daypart:

• Gender, Tempo, Intensity, Mood, Style controls

• Artist separation controls [global and individual artist]

• Prior-day horizontal title separation

• Artist blocks [multiple songs in-a-row by same artist]

• "Never-Violate" and "Preferred" rules

• Hour circulation rules

62

Automated Radio Programming

63

Social

Manual Automated

Non-Social 64

art of the mix

• Hand made playlists

• Mix art

• Web services

• Pre-crawled data at:
http://labrosa.ee.columbia.edu/projects/musicsim/aotm.html

65

ﬁql.com

• Browse / search for playlists

• Create a playlist:

• Search for artist / songs

• Add songs to a playlist

• Re-order the playlist

• Describe the playlist:

• title, description, tags

• Decorate the playlist

• Publish the playlist

66

Spotify
• Sharable playlists

• Collaborative playlists

• Many 3rd party playlist sites

69

Mix Enablers
mixcloud

70

Mix Enablers
mixcloud

• Free social networking platform organized around the
exchange of long form audio, principally [dance] music

• Provides a means for DJs (aspiring and professional) to
connect with the audience and into the Web of Things

70

Mix Enablers
mixlr

71

Mix Enablers
mixlr
• focused on adding social
features to centralized
multicasting

• supports live and recorded
(mixed and unmixed) streams

• social connectivity is web-
based, broadcaster is a native
application

• native app provides integration
with common DJ tools

72

setlist.fm
A wiki for
concert setlists

73

setlist.fm
A wiki for
concert setlists

They have an
API!

73

The Playlisting Dead pool

74

types. Further research is planned on how to allow user indexing of music assets.
Once a programme has been built it can be played immediately and is automatically
saved to the users profile for future retrieval. Programmes that are played more
than three times are awarded the top score of 5, even though the average rating of

Personal Radio
constituent items may be lower. Our theory is that a well chosen collection of
music has greater value than the sum of its constituent items. For one thing, there is
some work involved in putting together a programme so there is some value in
choosing something “off the shelf”. For another, a collection of music may contain
the difficult to quantify feature of “mood” which depends on the collected items
being played together. This feature is apparent where users amend their ratings for
individual items as they appear in different programmes. Figure 2 illustrates an
excerpt for the programme mellow and jazzy in which the user cchayes has rated

•
four out of the five shown items. If cchayes chooses mellow and jazzy again he will
An early collaborative filtering be shown his ratings for the individual items within the programme and he may
recast his vote. This facility is important because music taste does shift, and user
system profiles will have to move to reflect this. It is entirely probable that a user will
cease to become a recommender in one neighbourhood only to have moved to

•
another.
Users rated songs directly

• Playlists are built by finding
similar (via Pearson’s correlation
coefficient) users

• Playlists can, once built, be
streamed, named, shared and
modified Figure 2: a portion of play list entitled mellow and jazzy

• Order is either random or user
defined

Smart radio: Building music radio on the fly
Conor Hayes and Pádraig Cunningham 77

Equation 2: Pearson correlation coefficient
In equation 2, m refers to the number of items the two users have in comm
order to ensure that correlations are not being calculated over a small num

Personal Radio
common items a further weight is applied. With our current user popula
found it was necessary to have rated 20 items in common befor
recommendations were being made. Therefore, if a pair of users has less
items in common the correlation obtained by the Pearson measure is deva
m/20.

• An early collaborative filtering
system

• Users rated songs directly

• Playlists are built by finding
similar (via Pearson’s correlation
coefficient) users

• Playlists can, once built, be
streamed, named, shared and
modified

• Order is either random or user
defined
Figure 1: Naming a recently built play list

Smart radio: Building music radio on the fly
Conor Hayes and Pádraig Cunningham 77

comprises standard CD ripping and MP3 collection management recent vote winners (bottom left of figure 2).
software. Being connected to the Internet, the device also retrieves
from freedb.org and amazon.com, related information and images The main unit also serves numerous handheld clients (HP i
about the song, such as artist and album names and collaborative distributed on the tables throughout the bar (see figure 3).

Collaborative Choice
filtering information (e.g. “people who like this song also like these
JUKOLA traditional Jukebox, the nominated song is not guaranteed to be
artists”). The owner ofnumber of different an initial pool of musicplayed. Rather, it is subject to voting by other people in the public
Jukola is made up of a the space creates components which all and
organises it into different collections that can be activated according The interface also presents information about the song that is
afford different levels of control over the music choice. Music is space.
database on the main unit different currently playing (top left of figure 2) as well a short history of the
to the musical ambiencea appropriate for that space at that also times
stored as MP3 files in
of the day standard CD ripping and MP3 collection management recent vote winners (bottom left of figure 2).
comprises or week.
software. Being connected to the Internet, the device also retrieves
from freedb.org and amazon.com, related information and images The main unit also serves numerous handheld clients (HP iPAQs)
A public voting system
The main Jukola unit serves various different clients over a wireless
about the song, such as artist and album names and collaborative distributed on the tables throughout the bar (see figure 3).
network. The first of these is a 15-inch touch screen display that is
filtering information (e.g. “people who like this song also like these
situated in the public the space creates (see figure 1).of music and
artists”). The owner of part of the bar an initial pool
organises it into different collections that can be activated according
to the musical ambience appropriate for that space at different times
of the day or week.

The main Jukola unit serves various different clients over a wireless
network. The first of these is a 15-inch touch screen display that is
situated in the public part of the bar (see figure 1).

Figure 3. The handheld client used to vote for next s

The interface on the handheld client presents four candid
for the next song to be played. These candidate songs a
from the list of songs nominated on the public display as
Figure 1. Touch screen public display for random from the selected collection (the ratio of ra
Figure 3. The handheld client used to vote for next song.
nominating songs in the bar. nominated songs is dependent on number of songs
The interface on the handheldthe current song iscandidate songs
nominated). While client presents four playing, anyone i
The interface on the public display (see figure 2) essentially allows next song to be one of the handhelds can register drawnvote s
for the with access to played. These candidate songs are their
clientele to browse through the music collection and nominate thetouching on one of the on thecandidate songs. well asiPAQ a
from list of songs nominated public display as
four (the ratio of random to
Each
at
random from the selected collection
songs toFigure 1. Touch screen public display for
be played by pressing on them. vote per voting round - a voting round being the durati
nominating songs in the bar. nominated songs is dependent on number of songs currently
Jukola: democratic music choice in a public space nominated). While the current song is represented by a timeline at t
song currently playing and playing, anyone in the bar
K. O’Hara, M. Lipson, M. interface on the Jeffries, and P. Macer the display. A vote can be register their vote simply by78
The Jansen, A. Unger, H. public display (see figure 2) essentially allows with access to one of the handhelds can changed at any point during t

Collaborative Choice
activity. The same playlist also provides a vehicle by which songs
can be hyperlinked through to on-line vendors such as Amazon.com
(this draws on observations from earlier field work on lost impulses
use or immediately afterward
questions around their visit to th
the system as well as unpackin

Decentralized supply
whereby people hear songs in the environment they wish to buy but episodes of use. Where possib
then subsequently forget about them when an opportunity to elaborate on specific observatio
purchase arises [e.g.15, 16]. system. There were also op
comments to be collected when p
Short questionnaires were also u
the clientele who had used th
particular visit. After the trial, in
with the Watershed staff in orde
the system, their views on the m
café/bar, and the ways in which i
to the way they could manage t
Jukola web page were collecte
submitted via the web page.

The Watershed café bar
The Watershed offers various am
photographic dark rooms, conf
various exhibition rooms. As
amenities, the café bar is well
right with people visiting there
other amenities available.

Because of its status as a med
acquired somewhat of a repu
“intellectual” clientele. In actu
Figure 4. The web interface. diversity of people, including
people, families, individuals, an
The second
Jukola: democratic music choice in a public space key feature of the web page is a music upload read newspapers and books,79 mak
capability that allows the broader community to contribute to the
K. O’Hara, M. Lipson, M. Jansen, A. Unger, H. Jeffries, and P. Macer

have the music in the channel reflect the status of the members.
conducted 13
0 years old (7 The music played is broadcasted to each listener’s mobile device
upper secondary through a server. Everyone hears the same music as the other
study and our
a set of design
ype. Evaluating
Playlist Sharing
members currently listening. The listeners device displays
information of the current song and which user assigned it.

field test with a
ibed below. 6. PROTOTYPE
• Music should helpis implemented
Social Playlist convey
status informationaand
in Java with server-client
implicit presence client runs on
solution. The
Nokia S60 phone with 3G
connectivity. The server stores
• Music should help build listeners
information about the
tist or genre of interpersonal relationships
and their music selections.
Songs are stored on the server
music choice is
and broadcasted to the client
•
ng. While many A good individual listening
ation of status experience should be
devices at listening. The client
allows users to listen to the
mselves to using supported
channel and to change their
approach aim to
current activity or location.
c and everyday
oup. • Support smoothdisplays current
The client continuous
song title, artist and album
use together with the name of the
rsonal member which selected the
song. Figure 1 shows the Figure 1. Client interface for
and playlist:Roger Andersson Reimerinterface of ongoing relationships through collaborativeSocial Playlist prototype.
Social
topicenabling touch points and enriching the client.
KuanTing Liu and for the
mobile music listening
80

have the music in the channel reflect the status of the members.
conducted 13
0 years old (7 The music played is broadcasted to each listener’s mobile device
upper secondary through a server. Everyone hears the same music as the other
study and our
a set of design
ype. Evaluating
Playlist Sharing
members currently listening. The listeners device displays
information of the current song and which user assigned it.

field test with a
ibed below. 6. PROTOTYPE
Social Playlist is implemented
1. Members associate music from
theirin Java with a server-client
personal library to their
activities and locations runs on
solution. The client
Nokia S60 phone with 3G
connectivity. The server stores
2. For each new song, the system
information about the listeners
picks a random musicand a song
and their user selections.
tist or genre of
music choice is
fromSongs user’s currentthe server
that are stored on state
and broadcasted to the client
ng. While many
devices at listening. The client
ation of status Music is streamed to each
3. allows users to listen to the
mselves to using mobile device
channel and to change their
approach aim to
current activity or location.
c and everyday
oup. 4. The The client displays current
device displays the current
songsong which artist assigned it
and title, user and album
rsonal together with the name of the
member which selected the
song. Figure 1 shows the Figure 1. Client interface for
and playlist:Roger Andersson Reimerinterface of ongoing relationships through collaborativeSocial Playlist prototype.
Social
topicenabling touch points and enriching the client.
KuanTing Liu and for the
mobile music listening
80

Field Tested:
• Music should help convey
status information and
implicit presence

• Music should help build
interpersonal relationships

• A good individual listening
experience should be
supported

• Support smooth continuous
use

Social playlist: enabling touch points and enriching ongoing relationships through collaborative mobile music listening
KuanTing Liu and Roger Andersson Reimer 81

Field Tested:
"I am a weather guy. Happy music for
sunny days so to speak."
implicit presence

• Music should help build
interpersonal relationships

supported

use


Field Tested:
implicit presence

• Music should help build “I made her a CD because I can’t
interpersonal relationships stand her music.”

supported

use


Field Tested:
implicit presence


Participants report on hearing
experience should be between 30% - 50% “bad songs”.
supported

use


Field Tested:
implicit presence


Participants report on hearing
experience should be between 30% - 50% “bad songs”.
supported
At such occasions, they may turn off
• Support smooth continuous the service and switch to their own
use music library.


Implications

• Smooth integration with individual music listening to encourage
continuous use

• Allow ﬂexibility and cues to support self- expression and enable
touch points

• Support ongoing relationships

• Counterbalance experiences of bad songs and misinterpretations


Nearest Neighbors

84

Pure Content

• Uses MFCCs and ﬁnds N nearest neighbors
• Forms a graph with the all songs weighted
by distance
• Playlist is created by ﬁnding the shortest
weighted path covering N songs

Content-Based Playlist Generation: Exploratory Experiments
Beth Logan 85

song by the same artist or on the same album. Note that these results
give only an indication of performance. For example, several of our The results in this tabl
in the Same Artist an
genre categories overlap (e.g. and ) and songs from both

Pure Content
Table 2. This suggests
categories might still be perceived as relevant by a human user. in which labeling info
distance measure. Als
incorporated into play
3.2 Song Trajectory Playlists
The top part of Table 3 shows results for playlists formed from 4. CONCLUS
song trajectories. We show results for both variations discussed in We have investigated
The results in
playlists from a given
Section 2.1. The results show that the technique of tracing paths in the Same
though the song space gives worse results than the baseline. The previously published
Table 2. This
songs to a seed [2]. Th
second variation is somewhat better than the ﬁrst however. in which labe
through the distance
distance meas
feedback.
incorporated
We evaluated our tech
3.2 Song Trajectory Playlists varied 4.
styles. CON
Surprisi
The top part of Table 3 shows results for playlists formed from as well as simply cho
song trajectories. We show results for both variations discussed in We have inve
the playlist. We attrib
Section 2.1. The results show that the technique of tracing paths playlists from
measure. However,
though the song space gives worse results than the baseline. The added, previously pu
improvements
second variation is somewhat better than the ﬁrst however. a framework for a see
songs to incor
through the d
5. REFEREN
feedback.
[1] M. Alghoniemy a
playlist generation
We evaluated
[2] B. varied styles.
Logan and A.
function. Technic
as well as sim
3.3 Automatic Relevance Feedback oratory, June 2001
the playlist. W
Content-Based Playlist Generation: Exploratory Experiments
Beth Logan The second part of Table 3 shows results for automatic relevance [3] B. measure. 86Ho
Logan and A.

paper, whenever we refer to a music metadata vector, we mean a vector consisting of 7
categorical variables: genre, subgenre, style, mood, rhythm type, rhythm description, and
vocal code. This music metadata vector is assigned by editors to every track of a large

Metadata Models
corpus of music CDs. Sample values of these variables are shown in Table 1. Our kernel
function K(x, x ) thus computes the similarity between two metadata vectors correspond-
ing to two songs. The kernel only depends on whether the same slot in the two vectors are
the same or different. Speciﬁc details about the kernel function are described in section 3.2.
Metadata Field Example Values Number of
Values
Genre Jazz, Reggae, Hip-Hop 30
Subgenre Heavy Metal, I’m So Sad and Spaced Out 572
Style East Coast Rap, Gangsta Rap, West Coast Rap 890
Mood Dreamy, Fun, Angry 21
Rhythm Type Straight, Swing, Disco 10
Rhythm Description Frenetic, Funky, Lazy 13
Vocal Code Instrumental, Male, Female, Duet 6
Table 1: Music metadata ﬁelds, with some example values

3

Learning a Gaussian Process Prior for Automatically Generating Music Playlists
John C. Platt and Christopher J.C. Burges and Steven Swenson and Christopher Weare and Alice Zheng 87


Metadata Models
Values

• Use Gaussian Process Regression to create playlists based on seed
tracks 3



Metadata Models
Values

• Use Gaussian Process Regression to create playlists based on seed
tracks 3

• Using Kernel Meta-Training algorithm on albums to select the priors


most importantly, the kernel that came out of KMT is substantially better than the hand-
designed kernel, especially when the number of positive examples is 1–3. This matches the
hypothesis that KMT creates a good prior based on previous experience. This good prior
helps when the training set is extremely small in size. Third, the performance of KMT +

Metadata Models
GPR saturates very quickly with number of seed songs. This saturation is caused by the
fact that exact playlists are hard to predict: there are many appropriate songs that would be
valid in a test playlist, even if the user did not choose those songs. Thus, the quantitative
results shown in Table 2 are actually quite conservative.
Playlist 1 Playlist 2
Seed Eagles, The Sad Cafe Eagles, Life in the Fast Lane
1 Genesis, More Fool Me Eagles, Victim of Love
2 Bee Gees, Rest Your Love On Me Rolling Stones, Ruby Tuesday
3 Chicago, If You Leave Me Now Led Zeppelin, Communication Breakdown
4 Eagles, After The Thrill Is Gone Creedence Clearwater, Sweet Hitch-hiker
5 Cat Stevens, Wild World Beatles, Revolution
Table 3: Sample Playlists
To qualitatively test the playlist generator, we distributed a prototype version of it to a few
individuals in Microsoft Research. The feedback from use of the prototype has been very
• Use Gaussian Processof the playlist generator are shown in Table 3. In that table,
positive. Qualitative results Regression to create playlists based on seed
tracks
two different Eagles songs are selected as single seed songs, and the top 5 playlist songs
are shown. The seed song is always ﬁrst in the playlist and is not repeated. The seed song
on the left is softer and leads to a softer playlist, while the seed song on the right is harder
• Using Kernel a more hard rock play list.
rock and leads to Meta-Training algorithm on albums to select the priors

5 Conclusions
• Playlists are formed based on the maximum log likelihood from the
We have presented an algorithm, Kernel Meta-Training, which derives a kernel from a
selected seed song
set of meta-training functions that are related to the function that is being learned. KMT
permits the learning of functions from very few training points. We have applied KMT to
create AutoDJ, which is a system for automatically generating music playlists. However,

Metadata Models
Number of Seed Songs
Playlist Method 1 2 3 4 5 6 7 8 9
KMT + GPR 42.9 46.0 44.8 43.8 46.8 45.0 44.2 44.4 44.8
Hamming + GPR 32.7 39.2 39.8 39.6 41.3 40.0 39.5 38.4 39.8
Hamming + No GPR 32.7 39.0 39.6 40.2 42.6 41.4 41.5 41.7 43.2
Random Order 6.3 6.6 6.5 6.2 6.5 6.6 6.2 6.1 6.8
Table 2: R Scores for Different Playlist Methods. Boldface indicates best method with
statistical signiﬁcance level p < 0.05.
max

•
where Rj is the score from (11) if that playlist were perfect (i.e., all of the true playlist
Use Gaussian head of the RegressionR score of 100 indicates perfect prediction.
songs were at the Process list). Thus, an to create playlists based on seed
tracks for the 9 different experiments are shown in Table 2. A boldface result shows
The results
the best method based on pairwise Wilcoxon signed rank test with a signiﬁcance level of
0.05 (and a Bonferroni correction for 6 tests).
•
Usingare several Meta-TrainingTable 2. First, onof the experimental systems perform
There Kernel notable results in algorithm all albums to select the priors
much better than random, so they all capture some notion of playlist generation. This

•
Playlists are formed basedwent the of KMT isthe metadata schema. from the
is probably due to the work that
most importantly, the kernel that came out
on into designing substantially better than the hand-
maximum log likelihood Second, and
selected seedespecially when the number of positive examples is 1–3. This matches the
designed kernel, song
hypothesis that KMT creates a good prior based on previous experience. This good prior
helps when the training set is extremely small in size. Third, the performance of KMT +
87
John C. Platt and Christopher J.C. Burges and Steven Swenson and Christopher Weare and Alice Zheng

Traveling Sales Playlist?
b-based Combination
web-based genre classiﬁca-
yle detection on a set of 5
ining the predictions made
erfect overall prediction for
track similarity is linearly
similarity to obtain a new
gment an interface to music
rom the web. The interface
land landscape that places
their sound similarity. The
rtual environment. The ex-
ing terms on the landscape
tent in that region and the
provides semantic feedback

eneration
ration is treated as a net-
Figure 1: A screenshot of our Java applet “Trav-
n a start track and an end
eller’s Sound Player”.
algorithm ﬁnds a path (of
e Combining Audio-based Similarity with Web-based Data to Accelerate Automatic Music Playlist Generation
network satisfying user-
is labeled with Markus Schedl, and Gerhardwe incorporate web-based data to reduce the number of nec-
Peter Knees, Tim Pohle, a number Widmer 88

Traveling Sales Playlist?
• Using a combination of content-based song
and web-based artist similarity to generate a
distance matrix
• Approximation of TSP is used to ﬁnd ‘tours’
through the collection
• Tested on two collections of about 3000
tracks

Combining Audio-based Similarity with Web-based Data to Accelerate Automatic Music Playlist Generation
Peter Knees, Tim Pohle, Markus Schedl, and Gerhard Widmer 89

REGG
FOLK
CELT
METL
BOSS
ITAL
ITAL
REGG
PUNK
ACAP
FOLK
RAP
RAP

Now With Web Data
BLU
BLU
BOSS
JAZZ
ACAP
PUNK
ACJZ
ELEC
CELT

ELEC
ITAL
ACJZ
CELT
RAP
RAP
BOSS
ELEC
REGG
BLU
FOLK
BOSS
ACJZ
ITAL
CELT
JAZZ
PUNK
JAZZ
METL
BLU
FOLK
BLU
ITAL
RAP
BOSS
ACJZ
ACAP
JAZZ
ELEC
PUNK
CELT
PUNK
REGG
ACAP
METL
REGG
FOLK
ACAP
ELEC
FOLK
BLU
RAP
ACAP
RAP
ITAL
CELT
BOSS
ACAP
BOSS
JAZZ
BOSS
JAZZ
BLU
BLU
REGG
ACJZ
CELT
PUNK
ELEC
JAZZ
METL
ITAL
REGG
ELEC
CELT
ACJZ
FOLK
METL
ITAL
METL
PUNK
ACJZ
PUNK
RAP
ACAP
REGG
ELEC
ACJZ
FOLK
RAP
ITAL
FOLK
ITAL
REGG
METL
REGG
ACAP
ITAL
FOLK
CELT
REGG
BOSS
BLU
ACAP
ACAP
RAP
ACJZ
JAZZ
BLU
ELEC
BLU
BOSS
CELT
JAZZ
METL
PUNK
RAP
PUNK
ACJZ
METL
ELEC
PUNK 500 1000 1500 2000
CELT
Combining Audio-based Similarity with Web-based Data to Accelerate Automatic Music Playlist Generation
Peter Knees, METL
Tim Pohle, Markus Schedl, and Gerhard Widmer 90

Graph Methods
Dijkstra's algorithm
1. Assign to every node a distance value. Set it to zero for our initial
node and to infinity for all other nodes.
2. Mark all nodes as unvisited. Set initial node as current.
3. For current node, consider all its unvisited neighbors and calculate
their tentative distance (from the initial node).
4. When we are done considering all neighbors of the current node,
mark it as visited. A visited node will not be checked ever again; its
distance recorded now is final and minimal.
5. If all nodes have been visited, finish. Otherwise, set the unvisited
node with the smallest distance (from the initial node) as the next
"current node" and continue from step 3.

91

Graph Methods
∞
b
9 5
∞ 6 6
∞
2 4
11
∞
14 3
15
9 10

0 ∞
1 2
a 7

92

Graph Methods
∞
b
9 5
∞ 6 6
∞
2 4
11
∞
14 3
15
9 10

0 ?
1 2
a 7

92

Graph Methods
∞
b
9 5
∞ 6 6
∞
2 4
11
∞
14 3
15
9 10

0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
∞ 6 6
∞
2 4
11
?
14 3
15
9 10

0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
∞ 6 6
∞
2 4
11
9
14 3
15
9 10

0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
? 6 6
∞
2 4
11
9
14 3
15
9 10

0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
∞
2 4
11
9
14 3
15
9 10

0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
∞
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
∞
2 4
11
?
14 3
15
9 10
Out
0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
∞
2 4
9<10+7 11

14 3
15
9 10
Out
0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
?
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
22
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7

92

Graph Methods
∞
b
9 5
14 6
6
22
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
9 5
14 6
6
?
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
9 5
14 6
6
22>11+9
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
9 5
14 6
6
20
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
9 5
? 6 6
20
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
14>9+2 b
9 5
6 6
20
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
11 9 5
6 6
20
2 4
11
9
14 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
∞
b
11 9 5
6 6
20
2 4
11
9
14 Out 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
?
b
11 9 5
6 6
20
2 4
11
9
14 Out 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
20
b
11 9 5
6 6
20
2 4
11
9
14 Out 3
15
9 10
Out
0 7
1 2
a 7 Out

92

Graph Methods
minimum spanning tree

93

Graph Methods
min cut/max ﬂow
3
L M
1
2
1 4 P
3
2
4 5
N O
4
3 3
B E G I
2 1 1 1
2 1 2
A 1 4 F 1 4 K
3 3
2 2
2 1
C D H J
4 5

Q 3
2 R 4 6
1

S 2 T

Social Playlists and Bottleneck Measurements:
Exploiting Musician Social Graphs Using Content-Based Dissimilarity and Pairwise Maximum Flow Values 94
Fields, Ben and Jacobson, Kurt and Rhodes, Christophe and Casey, Michael

Graph-Based Path Finding
• A directed graph is created based on
the friend connections amongst artists
found on myspace
• The edges of this graph are weighted using
content-based similarity
• Playlists are constructed through the use of
the max ﬂow/min cut from a starting
to ending artist
Social Playlists and Bottleneck Measurements:
Exploiting Musician Social Graphs Using Content-Based Dissimilarity and Pairwise Maximum Flow Values 95
Fields, Ben and Jacobson, Kurt and Rhodes, Christophe and Casey, Michael

Our algorithm for computation of a playlist of length p (ex-
.
cluding start andfor computationplaylist of lengthsongs S ,
Our algorithm for computation of a aof a playlist of length p i(ex-
Our algorithm end song) for database of n p (ex-
cluding start and ends song)endingfor song Se consists songs Si ,
starting at start and and song) database of n of n ofi ,
cluding song S end for a at a database songs S the
Start-End Timbrel Paths
following steps: and and ending at Se consists of the
starting at song song Ss ending at song song Se consists of the
starting at Ss
following steps:steps:
following
1. for all i = 1, ..., n songs compute the divergences to
: 1. for the start 1, ...,song, calculate divergence from to
1. all i = song 1, KL (i,songs compute songdivergences
1.For all i = D ..., n s) and thethe divergences(i, e)
for every n songs compute end the DKL to
- the start song song(DKL (i, s)the endend (song DKL) e)
selectd%DKL (i, s) and )and divergenceKL (i, e) (i,
start and end
the start songs with greatest thesong D D (i, s)
2. find the
e KL
songs songs with find the d% songs with greatest
a
to find start song Ss ;greatest divergence DKLDKL (i, s)
2. the the
2. find the d% d% songs with greatest divergence (i, s)
e start start Ss (i,with thethe d% Se ; with greatest
to theto the song KL; find;to highestsongwith discard all
e) find end
divergence D song S the d% songssongs greatest
2. songs which are in both of these groups; keep remain-
Find d% songs s divergence from
- divergence (i, e) to the end end Ssong. ;Remove
divergencesong;DKL (i, e) to thesong songdiscard all all
DKL repeat against end e ; Se discard
startsongs further processing
ingwhich arefor both of these these groups; remain-
m
songssongs which are in both of groups; keep keep remain-
in
songs thatfurther processing sets.
appear in processing
both
ingfor songssongs for furthercompute a divergence ratio:
m allm=for ..., m songs
3. ing i 1,
3. all i = 1, ..., 1, ..., m songs compute remaining ratio:
Compute m songs compute a divergence ratio:
3. for for all i =
3. divergent ratio for a divergence
songs: DKL (i, s)
n R(i) = (3)
DKLKL (i,(i, s)
D D s) e)
(i,
KL
t =
R(i) R(i) = (3) (3)
4. compute step width for playlist: (i, e)
Playlist Generation Using Start and End Songs (i, e)
DKLDKL 97
Arthur Flexer, Dominik Schnitzer, Martin Gasser and Gerhard Widmer

3. for all i = 1, ..., m songs compute a divergence ratio:

Start-End Timbrel Paths R(i) =
DKL (i, s)
DKL (i, e)
(3)
ISMIR 2008 – Session 2a – Music Recommendat
4. Compute idealISMIR width:Session 2a – Music Recommend
4.
2008 –
compute step width for playlist:
step
5. compute p ideal positions (i.e. ideal divergence ratios)
R(s) − R(e)
ˆ
R(j), j = ideal step = (i.e. ideal divergence ratios) (4)
5. compute p 1, ...,positions p + 1
p:
ˆ
R(j), j = 1, ..., p :
5. Generate ideal positions jfor each song: (5)
ˆ
R(j) = R(s) + ∗ step
t
rt
ˆ
R(j) = R(s) + j ∗ step (5) ur
6. select the p real songs Sj that best match the ideal
eu
ˆ
divergence ratios songs j = that best match the ideal
R(j), S 1, ..., p :
6. select the p real j
6. Select ideal songs that 1, ..., pmatch the ideal: e
divergence ratios R(j), j = best :
ˆ
ˆ
Sj = arg min |R(j) − R(i)| (6)
i=1,...,m Table
ˆ
Sj = arg min |R(j) − R(i)| (6) (neare
The main part of our algorithm is the computation of di-
i=1,...,m Tabl
Playlist Generation Using Start and End Songs
Arthur Flexer, Dominik Schnitzer, Martin Gasser and Gerhard Widmer sults a
98

Evaluating S-E Paths
objective analysis

• The playlist should contain mostly songs
from genres A and B
• At the beginning of the playlist, most songs
should be from genre A, at the end from
genre B and from both genres in the middle

Arthur Flexer, Dominik Schnitzer, Martin Gasser and Gerhard Widmer 99

objective analysis
ISMIR 2008 – Session 2a – Music Recommendation and Organization

HiHo Regg Funk Elec Pop – Music Recommendation and Organization Funk Elec Pop Rock
ISMIR 2008 – Session 2a Rock HiHo Regg
Sec1 33 5 2 15 8 38 Sec1 26 7 2 20 7 38
Sec2 5 1 2 7 4 81 Sec2 6 1 2 7 4 80
Sec3 HiHo Regg Funk Elec Pop Rock
2 0 3 4 2 88 Sec3 HiHo Regg Funk Elec Pop Rock
3 0 2 4 2 88
Sec1 33 5 2 15 8 38 Sec1 26 7 2 20 7 38
Sec2 5 1 2 7 4 81 Sec2 6 1 2 7 4 80
Table 3. Distribution of songs across genres in playlists Table 6. Distribution of songs across genres in playlists
Sec3 2 0 3 4 2 88
starting at Hip Hop and ending at Rock. Results given for Sec3 3 0 2 4 2 88
starting at Reggae and ending at Rock. Results given for
first, middle and last section of playlists (Sec1 to Sec3). first, middle and last section of playlists (Sec1 to Sec3).
HiHo and Funk Elec Pop Rock
starting at Hip Hop Reggending at Rock. Results given for HiHo Regg Funk Elec Pop Rock
starting at Reggae and ending at Rock. Results given for
Sec1 30 5 2 35 8 19
first, middle and last section of playlists (Sec1 to Sec3). first, middle 19 last section of playlists (Sec1 to Sec3).
Sec1 and 3 8 28 13 29
Sec2 6 2 3 66 5 18 Sec2 17 4 4 20 19 36
Sec3 HiHo Regg Funk Elec Pop Rock
2 2 3 70 4 18 Sec3 HiHo Regg Funk Elec Pop Rock
12 3 4 22 16 42
Sec1 30 5 2 35 8 19 Sec1 19 3 8 28 13 29
Sec2 6 2 3 66 5 18 Sec2 17 4 4 20 19 36
starting at Hip Hop and ending at Electronic. 4Results given
Sec3 2 2 3 70 18 starting at Funk and ending at Pop. Results 16
Sec3 12 3 4 22 42
given for first,
for first, middle and last section of playlists (Sec1 to Sec3). middle and last section of playlists (Sec1 to Sec3).
starting at Hip Hop andStart and EndElectronic. Results given
Playlist Generation Using ending at Songs
starting at Funk and ending at Pop. Results given for first,
Thefirst, middle and last section and Gerhard Widmer(Sec1 to Sec3). Funk is confused with almost all other genres and genre
for start genresSchnitzer, Martin Gasser of playlists end genres are
Arthur Flexer, Dominik diminish quickly and the middle and last section of playlists (Sec1 to Sec3). 99
Pop strongly with genre Rock. As a result, the only visi-

subjective analysis

• How many outliers are in the playlist which
do not ﬁt the overall ﬂavour of the playlist?
• Is the order of songs in the playlist from the
start to the end song apparent?

Arthur Flexer, Dominik Schnitzer, Martin Gasser and Gerhard Widmer 100

subjective analysis
8 – Session 2a – Music Recommendation and Organization

refore, we look at only Genres # of order apparent
inations of our six gen- from to outliers yes somewhat no
b. 8). For each combi- HiHo Regg 4.7 x xx
randomly choose three HiHo Funk 1.7 xx x
s described in Sec. 4.1. HiHo Elec 1.3 xxx
. Our evaluator listened HiHo Pop 2.7 xx x
MS 1.2.10 - Cross plat- HiHo Rock 0 xxx
ld ﬁrst listen to the start Regg Funk 0.7 xx x
he songs in between in Regg Elec 1.3 xxx
allowed to freely move Regg Pop 1.3 xxx
moving back and forth
Regg Rock 0.3 xx x
e-listen to songs in the
Funk Elec 1.0 xx x
Funk Pop 1.7 xx x
was asked to answer the
Funk Rock 0 xx x
ightly connected to our
Elec Pop 0 xxx
.1:
Elec Rock 0 xx x
playlist which do not ﬁt Pop Rock 0 xxx
ist? average 1.1 71.1% 17.8% 11.1%
laylistFlexer, Dominik Schnitzer, to Gasser and Gerhard Widmer
Arthur from the start Martin 100

Playlist Similarity
• The co-occurrence of objects in an
authored stream can be used as a
proxy for object similarity
• This sort of similarity is especially effective
for the generation of playlists
• Employs the use of an undirected graph,
weighted by co-occurrence counts

Inferring similarity between music objects with application to playlist generation
R. Ragno and C.J.C. Burges and C. Herley 101

Playlist Similarity
ps: audio B

aborative 1
F
omparing 1
2 G
Metadata S
s such as 1 1

relies on 1 A 1
ns among E
2.
inferring 2
ural idea K
1 D

g what we
of infor- 1
“likeness” J
e play or-
ble infor- betweenFigure with application to playlist generation
Inferring similarity music objects
1: Graph representing the labeled stream 102
R. Ragno and C.J.C. Burges and C. Herley

A penalty function can directly bias a generated playlist
towards the original seed or given list (such as a partic- Lithium [Nirvana] : 0.0
ular radio station). This minimizes overall drift. Local Fall To Pieces [Velvet Revolver] 7.668
Nothin’ To Lose [Josh Gracin]
Tonight, Tonight [Smashing Pumpkins] 8.607

Playlist Similarity
song and the previous song is much tighter in style. This 12.712
low-probability choices for individual songs can be elim- Who’sHands Daddy [Toby Keith]
Your [Interpol] 13.695
can be extended indefinitely, of course, but it eventually Slow 12.712
inated by a cutoff on the number of times an arc must Want Fries With That [Tim McGraw] 8.607
degrades to matching only a particular radio station, Renegades Of Funk [Rage Against...] 10.127
be observed in the source data in order to be present
approximately. Hell YeahForget [Slipknot]
Before I [Montgomery Gentry] 14.214
7.355
in the graph (although that will bias heavily towards
The distribution of songs following the songs can also Awful, Beautiful Life [Darryl Worley] 12.607
The Kids Aren’t Alright [Offspring] 11.712

example playlists
songs that are frequently played).
be a mix of that following the current song and that Let Them Be Little [Billy Dean] [Killers] 9.542
All These Things That I’ve Done 13.777
following the previous song, with some discount factor Weapon [Matthew Good] 18.914
4. the previous song. This simpler approach still miti-
on EXPERIMENTS Again observeDoors the list stays entirely within the
that Down]
Kryptonite [3 11.127
gates the effects of choosing unlikely steps, and can also genre of [Three Days Grace]
Home Country music. Finally, starting with a Nir-
8.712
4.1 extended as far as desired.
be Examples Playlists vana song: [Godsmack]
Whatever 10.127
Wepenalty function few directly playlists to illustrate
A now present a can sample bias a generated playlist Colors [Crossfade] 7.097
the scheme. original seed or given listby a single partic-
towards the Each playlist is seeded (such as a song, Lithium [Nirvana] : 0.0
ular radio station). point.minimizes overall drift. Local
This The accompanying numbers
4.2 Fall To Pieces [Velvet Revolver]
Music Similarities 7.668
which is the starting
represent the distance from individualOur first example
low-probability choices for the seed. songs can be elim- Our randomTonight [Smashing Pumpkins] 12.712
Tonight, walk playlist generation induces a desir-
starts witha“Paperback Writer” by of times an arc must
inated by cutoff on the number the Beatles: able Slow Hands unpredictability. However to evaluate
variety and [Interpol] 12.712
be observed in the source data in order to be present our similarity measure, we alsoAgainst...]
Renegades Of Funk [Rage list the shortest path
10.127
in Paperback Writer [Beatles] bias heavily towards
the graph (although that will 0.0 songs for a I Forget of different seed songs. Note that
Before number [Slipknot] 7.355
songs that are frequently[Supertramp]
Breakfast In America played). 8.607 the resulting playlists Alright much more closely to the
The Kids Aren’t adhere [Offspring] 11.712
We’re An American Band [Grand Funk Rrd] 8.607 seed All These Things Thatwalk playlists given above:
song than the random I’ve Done [Killers] 9.542
Weapon [Matthew Good] 18.914
4. InEXPERIMENTS
The Dark [Billy Squier] 17.244
Hey Jude [Beatles] Down] 0.000
I Shot The Sheriff [Eric Clapton] 12 .192 Kryptonite [3 Doors 11.127
Fat Bottomed Girls [Queen] 16.335 Lady Madonna Days Grace]
Home [Three [Beatles] 7.515
8.712
4.1 Examples Playlists Stones]
Jumpin’ Jack Flash [Rolling 13.723 Lucy In The[Godsmack]Diamonds [Beatles]
Whatever Sky With 7.515
10.127
We now present aWeekend [Loverboy] to illustrate
Working For The few sample playlists 15.251 Peace Of[Crossfade]
Colors Mind [Boston] 7.737
7.097
theDream Weaver [Gary Wright]
scheme. Each playlist is seeded by a single 15.520 song, (Just Like) Starting Over [John Lennon] 7.737
4.2 Music Similarities
Saturday In The Park [Chicago] 8.000
which is the starting Spirit! The accompanying numbers
Smells Like Teen point. [Nirvana] 15.735
representStopdistance from the seed. Our first example
Can’t the [Red Hot Chili Peppers] 16.732 Shine It All Around [Robert Plant] induces a desir-
Our random walk playlist generation 8.000
starts with “Paperback Writer” by the Beatles: 19.256
Still Waiting [Sum 41] Holiday [Green Day] 8.000
able variety and unpredictability. However to evaluate
Grave Digger [Dave Matthews] 20. 665 our similarity measure, we also list Brothers] 8.000
Rock And Roll Heaven [Righteous the shortest path
Paperback Writer [Beatles] 0.0 songs for a number of different seed songs. Note that
Note that the list America [Supertramp]category of mu-
Breakfast In stays within the broad 8.607 the resultingTo Hell [AC/DC] 0.000
Highway playlists adhere much more closely to the
sic that could American Band [Grand example it never
We’re An be considered close: for Funk Rrd] 8.607 seed song than [Foo random walk playlists given above:
Best Of You the Fighters] 6.252
strays into Jazz, Country, Hip Hop or Punk. Our next
In The Dark [Billy Squier] 17.244 Remedy [Seether] 6.362
example is The Sheriff [Eric Clapton] Your Man” 12 .192
I Shot a Country song “Stand by by Right Here [Staind]
Inferring similarity between music objects with application to playlist generation
Hey Jude [Beatles] 6.362 0.000
R. Ragno and C.J.C. Burges and C. Herley 103
HolidayMadonna [Beatles]
Lady [Green Day] 6.362 7.515

which is the starting[3 Doors Down]
Kryptonite point. The accompanying numbers 11.127
represent the distanceDays Grace]
Home [Three from the seed. Our first example 8.712 Our random walk playlist generation induces a desir-
starts with “Paperback Writer” by the Beatles:
Whatever [Godsmack] 10.127 able variety and unpredictability. However to evaluate

Playlist Similarity
e Colors [Crossfade] 7.097 our similarity measure, we also list the shortest path
, Paperback Writer [Beatles] 0.0 songs for a number of different seed songs. Note that
4.2 Music Similarities the resulting playlists adhere much more closely to the
s Breakfast In America [Supertramp] 8.607
e We’re An American Band [Grand Funk induces8.607
Our random walk playlist generation Rrd] a desir- seed song than the random walk playlists given above:

example similarities
Inable variety[Billyunpredictability. However to evaluate
The Dark and Squier]
I Shotsimilarity measure, we also list the shortest .192
our
The Sheriff [Eric Clapton]
songs for a number of different seed songs. Note that
Fat Bottomed Girls [Queen]
17.244
12 path
16.335
Hey Jude [Beatles]
Lady Madonna [Beatles]
0.000
7.515
07 Jumpin’ Jack Flash [Rolling Stones] more closely to the
the resulting playlists adhere much
13.723 Lucy In The Sky With Diamonds [Beatles] 7.515
07 Working For Thethe random walk playlists given above:
seed song than
Weekend [Loverboy] 15.251 Peace Of Mind [Boston] 7.737
244 Dream Weaver [Gary Wright] 15.520 (Just Like) Starting Over [John Lennon] 7.737
192 Smells Like Teen[Beatles]
Hey Jude 0.000 Saturday In The Park [Chicago] 8.000
Spirit! [Nirvana]
Lady Madonna [Beatles] 15.735
7.515
335 Shine It All Around [Robert Plant] 8.000
Can’tLucy In TheHot Chili Peppers] [Beatles] 16.732
Stop [Red Sky With Diamonds 7.515
723 Holiday [Green Day] 8.000
Still Waiting [Sum 41]
Peace Of Mind [Boston] 19.256
7.737
251 Rock And Roll Heaven [Righteous Brothers] 8.000
Grave Digger [Dave Matthews][John Lennon] 20. 7.737
(Just Like) Starting Over 665
520
735 Saturday In The Park [Chicago] 8.000
Note that the list All Around [Robert Plant]
Shine It stays within the broad category of mu- 8.000 Highway To Hell [AC/DC] 0.000
732
sic that could be[Green Day] close: for example it never
Holiday considered 8.000 Best Of You [Foo Fighters] 6.252
256
strays into Jazz, Country, Hip Hop or Punk. Our next
Rock And Roll Heaven [Righteous Brothers] 8.000 Remedy [Seether] 6.362
665
example is a Country song “Stand by Your Man” by Right Here [Staind] 6.362
-Tammy Wynette: To Hell [AC/DC]
Highway 0.000 Holiday [Green Day] 6.362
r Best Of You [Foo Fighters] 6.252 Be Yourself [Audioslave] 6.5 58
t Stand By Your Man [Tammy Wynette] 6.362
Remedy [Seether] 0.0 The Hand That Feeds [Nine Inch Nail s] 6.584
y Chrome [Trace Adkins]
Right Here [Staind] 8.607
6.362 B.Y.O.B. [System Of A Down] 6.754
Stay With Me (Brass Bed) [Josh Gracin]
Holiday [Green Day] 8.607
6.362 Happy? [Mudvayne] 6.847
WhiskeyYourself [Audioslave]
Be Girl [Toby Keith] 14.162
6.5 58 Shine It All Around [Robert Plant] 6.982
ClassThe Hand [Lonestar] [Nine Inch Nail s] 6.584
Reunion That Feeds 13.965
07 My Sister [Reba McEntire] Down]
B.Y.O.B. [System Of A 6.754
12.650 Stand By Your Man [Tammy Wynette] 0.000
Happy? [Mudvayne]
07 Could Have Fooled Me [Adam Gregory] 6.847
12.777 You’ll Be There [George Strait] 5.800
162 Shine It All Around [Robert Plant] 6.982
965
650Inferring similarityBy Your Man [Tammy Wynette] 0.000
Stand between music objects with application to playlist generation
R. Ragno and C.J.C. Burges and C. Herley 78 104
777 You’ll Be There [George Strait] 5.800

Playlist Steering

• Create a timbrel features
• Create the space using tuple and triple n-
gram sequences from playlist logs
• Generate playlists via Tag Steering

Steerable Playlist Generation by Learning Song Similarity from Radio Station Playlists
Maillet, François and Eck, Douglas and Desjardins, Guillaume and Lamere, Paul 105

Playlist Steering
1. Select a seed track
2. Threshold transition matrix to generate set of possible
next tracks
3. User creates a tag cloud, assigning weights to any of 360
tags
4. Autotagger creates tag cloud for all candidate tracks
selected in (2). Cosine distance is taken between the
user’s tag cloud and each song’s.
5. The track with the minimum cosine distance from seed is
played

Steerable Playlist Generation by Learning Song Similarity from Radio Station Playlists
Maillet, François and Eck, Douglas and Desjardins, Guillaume and Lamere, Paul 106

milarity model is used to compute transitional
betweenRetrieval Conference (ISMIR(with
mation the seed song and all other ones 2009) Soft tag cloud
songs having higher transition probabilities), Viva la Vida by Coldplay
the top ϕ, or thresholding at a certain transi- Wish You Were Here by Pink Floyd

Playlist Steering
ty ρ. Let T be thethe following playlists are seeded with the
Table 4. Both group of these top songs:
song Clumsy by Our Lady Peace. To give a clear point
T reference, we use sthe )tag clouds of actual songs as the
of = arg maxϕ M(t , ti (1)
Peaceful, Easy Feeling by Eagles
With or Without You by U2
One by U2
ti ∈Tts
steerable cloud. The soft tag cloud is made up of the tags Fields Of Gold by Sting
er is then invited to create a tag cloud CUhard tag cloud with Every Breath You Take by The Police
for Imagine by John Lennon and the by
ghts to anyfor Hypnotize byin the system. In
the tags of the 360 tags System of a Down. Gold Dust Woman by Fleetwood Mac
loud is personalized to represent the mood or Enjoy The Silence by Depeche Mode
the user would like to hear. tag cloud the
Soft The higher Hard tag cloud
articular tag, the more impact it by Coldplay
Viva la Vida will have on All I Want by Staind
of the next song. Wish You Were Here by Pink Floyd Re-Education (Through Labor) by Rise Against
ger is used to generate a tag cloud Ctj by Eagles
Peaceful, Easy Feeling for all Hammerhead by The Offspring
. The cosine distance (cd(·)) betweenby U2
With or Without You these The Kill by 30 Seconds To Mars
d CU is used to ﬁnd the songOne by U2matches
that best When You Were Young by The Killers
Fields Of Gold by Sting
musical context the user described with his or Hypnotize by System of a Down
Every Breath You Take by The Police Breath by Breaking Benjamin
tmin = arg min cd(CU , Ctj ) by Fleetwood Mac
Gold Dust Woman (2) My Hero by Foo Fighters
tj ∈T Enjoy The Silence by Depeche Mode Turn The Page by Metallica
k tmin is selected to playHard tag cloud sys- next. Since the
arent, we can tell the user Iwe choseStaind
All Want by the song
songs is more important to us than the relative global place-
it has a certain transition probabilityby Rise Against
Re-Education (Through Labor) from
ment of, e.g., jazz with respect to classical). We have over-
g but also because its tag cloud The Offspring
Hammerhead by overlapped
laid the trajectory of the two playlists in Table 4 to illustrate
particular way. The Kill bycan Seconds To Mars
The user 30 then go back
their divergence.
heSteerable Playlist U to inﬂuence how subsequent Radio Station Playlists
tag cloud C Generation by Learning Song Similarity from
When You Were Young by The Killers
107
selected.
Maillet, François and Eck, Douglas and Desjardins, Guillaume and Lamere, Paul
Hypnotize by System of a Down

Playlist Steering Oral Session 4: Music Recommendation

[5] B
ci
in
si
ce
In
[6] P.
R
N
[7] T.
to
tic
Figure 1. Part of the 2-d representation of the track-to-
Steerable Playlist Generation by Learning Song Similarity from Radio Station Playlists M
107
Maillet, François and Eck, Douglas and Desjardins, Guillaume and Lamere, Paul

Scaling up playlist generation
• Building playlists involves satisfying constraints. e.g.

• Global constraints: No duplicate songs, No
consecutive artists, tempo between 120 and 130
BPM

• Ordering constraints: no consecutive artists,
DMCA rules

• Sorting constraints: ordered by danceability and
loudness

• Playlist length: 15 songs, 32 minutes, < 20mb

• Finite constraint satisfaction problem. It’s NP-HARD

109

General Approach

• Playlist is a sequence of songs: S1, S2 ... Sn
drawn from a large pool of songs

• Cost(Sn, C) is how well song S at position N
satisﬁes constraint C

• Cost(Sn) is total cost for song S at position N for
all constraints

• Cost(P) is total cost of all songs in the Playlist

• Goal: Find S1, ... Sn that minimizes Cost(P)

110

Scaling up playlist generation
Generate random playlist

while Cost(P) > threshold:
Calculate Cost(Sn) for each song
find max( Cost(sN) ) that is not Tabu
find best possible replacement

worst variables for which no value can be found to decrease the total cost
are labelled as Tabu for a given number of iterations.
Typical runtime: 1.4 seconds for 10 song playlist from a pool of 20,000
songs with 10 constraints

SCALING UP MUSIC PLAYLIST GENERATION
Jean-Julien Aucouturier, Francois Pachet 111

Case-based sequential ordering
of songs for playlist recommendation

Case-based Sequential Ordering of Songs for Playlist Recommendation⋆
Claudio Baccigalupo and Enric Plaza 112

Fast Generation of Optimal Music Playlists
using Local Search
• Simulated Annealing

• Heuristic Improvements

• Song domain reduction

• Two level search:

• 1: Replace, Insert Delete

• 2: Swap

• Partial constraint voting

Typical runtime: 2 seconds
for 14 song playlist with 15
constraints from a pool of
2,000 songs
Fast Generation of Optimal Music Playlists using Local Search
Steffen Pauws, Wim Verhaegh, Mark Vossen 113

Echo Nest Playlister
world of songs

• Start with millions of
songs
playlist rules
initial song selection
song pool • Apply global
constraints to create
smaller song pool
(1K to 10K songs)

• Use constraint engine
song constraint
satisfaction engine
to ﬁnd best playlist:

• Beam search
ﬁnal playlist generation
• Adaptive search
populate with data • Populate with data

114

Group Playlisting

• Group Playlisting:

• Radio, Clubs, Ofﬁces, Health clubs, The Web

• Group playlisting challenges

• Varying and conﬂicting music tastes

• Different levels of assertiveness

• Traditional

• Dictator, Compromise, Random, opt-out

129

Group Cost Functions

• New cost functions for group playlisting: social
cost function:

• Average happiness - group vote of members

• Maximum happiness - vote of the happiest group
member

• Minimum misery - vote of the least happy

Group Recommending: A methodological Approach based on Bayesian Networks
Luis M. de Campos, Juan M. Ferna ́ndez-Luna, Juan F. Huete, Miguel A. Rueda-Morales 130

Group costs
Ben Paul Tom Avg Max Min

2 10 1 4.33 10 1

4 3 3 3.33 4 3

6 2 7 5 6 2

131

Flytrap

• Uses simple voting mechanism - ‘average happiness’

• Each listener agent votes:

• Artist previously listened == high votes

• Genre previous listened == positive vote

• Songs with more votes have higher probability
of being played

• Never play 2 songs by same artist in a row

• Loose coherence of genre across tracks

Flytrap: Intelligent Group Music Recommendation
Andrew Crossen, Jay Budzik, and Kristian J. Hammond
132

Flycasting
1. Translate the request histories of all requesters
into ratings for artists.
2. Predict ratings for each artist that a requester has
never requested.
3. Determine what artists are the most popular
among the listening audience.
4. Determine what artists are similar to the ﬁnal
artist on the playlist.
5. Select a song to play that is performed by an artist
that is both popular among the listening requesters
and similar to the artist that precedes it.

Flycasting: On the Fly Broadcasting
James C. French and David B. Hauver 133

How to Combine Diﬀerent Individual Preferences

e goal of the Reuse Process is to combine diﬀerent individual
preferences into a global group ranking of the candidate songs

I Spy (Pulp) Ex.: three listeners have diverging individual preferences
retrieved over which candidate song to play after I Spy (Pulp)
candidates

Lazy (Suede) 0.9 0 0.6 ?

Go (Moby) 0 -1 0.9 ?

Uno (Muse) -0.7 -0.3 1 ?

Drive (R.E.M.) 0.2 0.2 0.2 ?

A Case-Based Song Scheduler
for Group Customised Radio
Claudio Baccigalupo – Enric Plaza 134


1. To avoid misery, any candidate song that is hated by some
listener automatically gets the lowest group preference degree

I Spy (Pulp)
retrieved
candidates

Lazy (Suede) 0.9 0 0.6 ?

Go (Moby) 0 -1 0.9 -1

Uno (Muse) -0.7 -0.3 1 ?

Drive (R.E.M.) 0.2 0.2 0.2 ?



2. To ensure fairness, the group preference degree of the remaining
candidates equals to the average of the individual preferences

I Spy (Pulp)
retrieved
candidates

Lazy (Suede) 0.9 0 0.6 0.75

Go (Moby) 0 -1 0.9 -1

Uno (Muse) -0.7 -0.3 1 0

Drive (R.E.M.) 0.2 0.2 0.2 0.2



3. To guarantee individual satisfactions, listeners whose preferred
song was not selected in this turn are to be favoured next

Lazy (Suede) 0.9 0 0.6 0.75

Go (Moby) 0 -1 0.9 -1

Uno (Muse) -0.7 -0.3 1 0

Drive (R.E.M.) 0.2 0.2 0.2 0.2

satisfied not satisfied not satisfied



4. e satisfaction degree of a listener for previous songs changes
her weight in the calculation of the average group preference

Lazy (Suede)
retrieved
candidates

Loser (Beck) 0.2 ! 1 0.6 ! 0 0.9 ! -1 -1

Song 2 (Blur) 0.2 ! 0 0.6 ! -0.3 0.9 ! 1 0.24

Flower (Eels) 0.2 ! -0.7 0.6 ! 0.8 0.9 ! 1 0.41

Joga (Björk) 0.2 ! 1 0.6 ! 0.6 0.9 ! -0.2 0.13

satisfied not satisfied not satisfied


Beat-matching and cross-fading

• Select songs with similar tempos

• Select transition location

• Similar rhythmic pattern

• Speciﬁc sections (last 30 seconds of song 1 and
ﬁrst 30 seconds of song 2)

• Align their beats over the course of a transition

• Cross-fade the volumes

Creating Music by Listening
by Tristan Jehan
140

First, ﬁnd the beats

by Tristan Jehan
141

Time scaling

Creating Music
by Listening
by Tristan Jehan
142

Beat-matching and cross-fading

by Tristan Jehan
143

Some Examples

Rihanna (122 bpm) (95 bpm) Gotan Project

144

Some Examples

Bob Marley to Bob Marley


144

Some Examples

Sade to Sting


144

Some Examples

Sade to Sting
April March to April March


144

Direct Listening Tests
hypotheses
1. Playlists compiled by PATS contain more
preferred songs than randomly assembled
playlists, irrespective of a given context-of-
use.
2. Similarly, PATS playlists are rated higher than
randomly assembled playlists, irrespective of a
given context-of-use.

PATS: Realization and User Evaluation of an Automatic Playlist Generator
Steffen Pauws and Berry Eggen 147

hypotheses
3. Successive playlists compiled by PATS
contain an increasing number of preferred
songs.
4. Similarly, successive PATS playlists are
successively rated higher.
5. Successive playlists compiled by PATS
contain more distinct and preferred songs
than randomly assembled playlists.


set-up
• Three measures: precision, coverage and rating score

• 20 participants (17m, 3f), 8 sessions over 4 days per
participant

• User selects a song, given a context (4 playlist per
context)

• A PATS playlist and a random playlist are generated
(11 songs each, 1 minute excerpts)

• Judgements expressed per song, ratings per playlist


time.
education. Sixteen participants played a musical instrument.

erpts of jazz
o) from 100
. The music
3.4 Results
Playlists contained 11 songs from which one was selected by the
participant. This song was excluded from the data as this song was
tyles cover a
yle contained
consider for analysis. results
not determined by the system, leaving 10 songs per playlist to

ess and sound 3.4.1 Precision
udgment. The The results for the precision measure are shown in Figure 3.
workstation,
1 B personal
m).
of a 17-inch
experimental
eferred level.
sitioned at a

nute excerpts)
d pre-defined
ments of the
Figure 3. Mean precision (and standard error) of the playlists
he songs were
in different contexts-of-use. The left-hand panel (a) shows
e which song
mean precision for both playlist generators (PATS and
he process of
random) in the ‘soft music’ context-of-use. The right-hand
s freely in any
panel (b) shows mean precision for both generators in the
ressed. There Eggen
Steffen Pauws and Berry 150

in mean precision between the fourth PATS playlist and mean
precision of preceding PATS playlists in contrast to randomly

assembled playlists (F(1,19) = 8.935, p < 0.01). In other words,
each fourth PATS playlist contained more preferred songs than the
preceding three PATS playlists (mean precision of fourth PATS

results
session: 0.76; mean precision of the first three PATS sessions:
0.67). No other effects were found to be significant. Figure
playlists in
3.4.2 Coverage shows me
The results for the coverage measure are shown in Figure 4. random)
panel (b) s

A MANOVA
(2), context-
subject ind
variable. A
found (F(1,1
were rated h
score: 7.3 (
playlists can
assembled p
effect for co
Figure 4. Mean coverage (and standard error) of the playlists Playlists for
in different contexts-of-use. Recall that coverage is a (mean rating
cumulative measure. The left-hand panel (a) shows mean significant e
coverage for both playlist generators (PATS and random) in
the ‘soft music’ context-of-use. The right-hand panel (b) shows
Steffen Pauws and Berry Eggen
3.4.4 151
Inte

of an Automatic Playlist Generator

nerator was
not already contained in earlier playlists. For comparison, the
1). Playlists
an randomly
ATS), 0.45
to be significant. results
random approach added four songs. No other effects were found

ound to be 3.4.3 Rating score
or the ‘soft The results for the rating score are shown in Figure 5.
ongs (mean
n interaction
t significant
test, it was
5). Further
nt difference
st and mean
o randomly
other words,
ngs than the
ourth PATS
TS sessions:
Figure 5. Mean rating score (and standard error) of the
playlists in different contexts-of-use. The left-hand panel (a)
shows mean rating for both playlist generators (PATS and
ure 4. random) in the ‘soft music’ context-of-use. The right-hand
PATS: Realization and User Evaluation(b) shows mean rating score for both generators in the
panel of an Automatic Playlist Generator

Skip-Based Listening Tests
basics
• Evaluation integrated into system
• Assumptions:
1. a seed song is given
2. a skip button is available and easily
accessible to the user
3. a lazy user who is willing to sacriﬁce
quality for time

Dynamic Playlist Generation Based on Skipping Behavior
Elias Pampalk and T. Pohle and G. Widmer 153

use cases
1. The user wants to listen to songs that are
similar to the seed song
2. Same as (1) but with a dislike of an arbitrary
artist for a subjective reason (eg taste)
3. The user’s preference changes over time.
Speciﬁcally, in a 20 song playlist, the ﬁrst 5
songs from genre A, the middle 10 from either
genre A or B, last 5 songs from genre B.


skips in UC1

Artists/Genre Tracks/Genre
Genres Artists Tracks Min Max Min Max
istance
22 103 2522 3 6 45 259
to the
date to Table 1: Statistics of the music collection.
a . If S
he best Heuristic Min Median Mean Max
UC-1 A 0 37.0 133.0 2053
B 0 30.0 164.4 2152
C 0 14.0 91.0 1298
D 0 11.0 23.9 425
UC-2
eElias Pampalk and T. Pohle and G. Widmer
user A 0 52.0 174.0 2230 155

D 0 11.0 23.9 425
that the user UC-2 A 0 52.0 174.0 2230

h is approxi- B 0 36.0 241.1 2502
C 0 17.0 116.9 1661
counted until D 0 15.0 32.9 453
(UC) are the

imilar to the
skips in UC1
Table 2: Number of skips for UC-1 and UC-2.

ty with genre 10
ed’s genre is

Mean Skips
5
music but dis-
asons such as
0
ame approach 1 5 10 15 20
d’s genre (not Playlist Position
ected. Every (a) Heuristic A
the unwanted
2
Mean Skips

me. We mea-
he seed song 1
o prefer. The
genre A. The 0
1 5 10 15 20
A or B. The Playlist Position
manually se-
e list of pairs (b) Heuristic D
C-2Pampalk and T. Pohle and G. Widmer on Skipping Behavior
Dynamic Playlist Generation Based
Elias it is pos- 155

GenresSkip-Based Listening Tests
Artists Tracks
Artists/Genre
Min Max
Tracks/Genre
Min Max
22
UC1 and UC2 skips
103 2522 3 6 45 259

Table 1: Statistics of the music collection.
Heuristic Min Median Mean Max
UC-1 A 0 37.0 133.0 2053
B 0 30.0 164.4 2152
C 0 14.0 91.0 1298
D 0 11.0 23.9 425
UC-2 A 0 52.0 174.0 2230
B 0 36.0 241.1 2502
C 0 17.0 116.9 1661
D 0 15.0 32.9 453

Table 2: Number of skips for UC-1 and UC-2.

UC3 skips

Heuristic A Heuristic B Heuristic C Heuristic D
Start Goto Median Mean Median Mean Median Mean Median Mean
Euro-Dance Trance 69.0 171.4 36.0 64.9 41.0 69.0 20.0 28.3
Trance Euro-Dance 66.0 149.1 24.0 79.1 6.5 44.4 4.5 8.8
German Hip Hop Hard Core Rap 33.0 61.9 32.0 45.6 31.0 40.7 23.0 28.1
Hard Core Rap German Hip Hop 21.5 32.2 18.0 51.9 16.0 24.2 14.0 16.1
Heavy Metal/Thrash Death Metal 98.5 146.4 54.0 92.5 58.0 61.1 28.0 28.4
Death Metal Heavy Metal/Thrash 14.0 69.2 16.0 53.7 3.0 55.5 3.0 25.7
Bossa Nova Jazz Guitar 68.5 228.1 32.0 118.7 54.0 61.1 22.0 21.3
Jazz Guitar Bossa Nova 21.0 26.7 22.0 21.5 9.0 10.5 6.0 6.2
Jazz Guitar Jazz 116.0 111.3 53.0 75.7 45.0 74.0 18.5 27.3
Jazz Jazz Guitar 512.5 717.0 1286.0 1279.5 311.0 310.8 29.0 41.3
A Cappella Death Metal 1235.0 1230.5 1523.0 1509.9 684.0 676.5 271.0 297
Death Metal A Cappella 1688.0 1647.2 1696.0 1653.9 1186.0 1187.3 350.0 309.2

Table 3: Number of skips for UC-3.

fail (e.g. electronic or downtempo). However, some of the playlist) due to the small number of artist per genre.
failures make sense. For example, before 20 pieces from The heuristic depends most of all on the similarity
electronic are played, in average almost 18 pieces from measure. Any improvements would lead to fewer skips.
downtempo are proposed. However, implementing memory effects (to forget past 157
Elias Pampalk and T. Pohle and G. Widmer

Dynamic Heuristics
• Last.fm Radio logs are used to analyze and
evaluate several heuristics for dynamic playlists

• This is done through the treatment of playlists as
fuzzy sets

• Work shows that one heuristic work best given
inconsistent rejects while another performs
best given inconsistent accepts and third
performs equally in either environment.

Evaluating and Analysing Dynamic Playlist Generation Heuristics Using Radio Logs and Fuzzy Set Theory
Bosteels, Klaas and Pampalk, Elias and Kerre, Etienne E. 158

Dynamic Heuristics and P
Oral Session 4: Music Recommendation

endation and Playlist Generation
(a) dataset 1 (b) dataset 3 (c) dataset 5

Figure 6. Two-dimensional histograms that illustrate how the 9 generated d
inconsistent accepts to a high level of inconsistent rejects.

50 50 50 50
aset 5 (d) dataset 7 (e) dataset 9

he 9 generated datasets gradually move from a high level of
40 40 40 40

30
Bosteels, Klaas and Pampalk, Elias and Kerre, Etienne E. 30 30 159
30

Dynamic Heuristics
(a) dataset 1 (b) dataset 3 (c) dataset 5 (d) dataset 7

Figure 6. Two-dimensional histograms that illustrate how the 9 generated datasets gradua
inconsistent accepts to a high level of inconsistent rejects.
: Music Recommendation and Playlist Generation
50 50 50 50

40 40 40 40

30 30 30 30

(c) dataset 5 (d) dataset 7 (e) dataset 9
20 20 20 20
2 4 6 8 2 4 6 8 2 4 6 8 2 4 6
hat illustrate how the 9 generated datasets gradually move from a high level of
(a) ISM (b) ISP (c) ISL = ITL (d) ITP
nsistent rejects.
Figure 7. Results of the additional evaluations for Ha (- -), Hb (–), and Hc (-·-). The num
I I

50 are dataset identiﬁers, while the vertical axis shows failure rate percentages.
50 50

40 40 IS L 40I
results described in [8], and Hc L perform at least
S
Ha
30 as well as all other instances of Ha and Hc , respectively.
30
I
30
I

20 20 20
8 2 4 6 8 2 4
2 4 66 8 8
7. CONCLUSION AND FUTURE WORK (a) inconsistent accepts
(c) ISL = ITL (d) ITP (e) ITM
The mathematical apparatus from the theory of fuzzy sets Figure 8. Categorization o
Bosteels, Klaas and
I
(- -), Hprovesandbe I (-·-). The numbers along dynamic playlist
tions for HaPampalk, Elias and Kerre, Etienne E.Hcvery convenient for deﬁningthe horizontal axis
b (–), to grained two-dimensional hi
160

Measuring Distance

We can measure the distance between
sequences of tracks using the same
methods we use to measure the distance
between frames within tracks.

Using Song Social Tags and Topic Models to Describe and Compare Playlists
Ben Fields, Christophe Rhodes and Mark d'Inverno 162

Measuring Distance
• Topic Modeled Tag Clouds used as a song-
level feature
• Sequences of these low dimensional
features can then be examined
• The ﬁtness of this pseudo-metric space is
examined through patterns in radio playlist
logs


Measuring Distance
gather tags for all songs

create LDA model describing
topic distributions

infer topic mixtures for all
songs

create vector database
of playlists


Measuring Distance


An evaluation of
various playlisting
services

Objective Evaluation

166

Some playlist stats
Playlist stats
art of the
Source Radio Paradise Musicmobs Pandora
mix

Playlists 45,283 1,736 29,164 94

Unique Artists 1,971 19,113 48,169 556

Unique Tracks 6,325 93,931 218,261 908

Average Length 4.3 100 20 11

% with duplicate
0.3% 79% 49% 48%
artist
% with
0.3% 60% 20% 5%
consecutive artists

Pandora playlist stats based on listening on 44 separate ‘stations’
167

Objective evaluation
Tag diversity
Playlist Tag Diversity
Source Tag Diversity Random
MusicMobs 0.29 / 0.18 0.51 / 0.13
Pandora 0.44 / 0.20 0.64 / 0.19
Art of the mix 0.48 / 0.17 0.61 / 0.11
Radio Paradise 0.75 / 0.13 0.75 / 0.13

Tag Diversity: unique artist tags vs. total artist tags
168

Radio Paradise diversity examples
Low Diversity Playlists
Artist Track Tags
Sun Volt Live Free Alt-country, americana, rock, country, folk, indie

indie, folk, singer-songwriter, americana, Alt-
Sun Kil Moon Gentle Moon country, alternative

folk, singer-songwriter, female vocalists, indie,
ANi DiFranco Angry Any More alternative, rock

Handcuffed to a fence in Alt-country, singer-songwriter, americana, folk,
Jim White
Mississippi indie, country

folk, female vocalists, singer-songwriter, indie,
Jess Klein Soda Water acoustic, girls with guitars

Diversity: 0.367
11 unique tags out of 30 169

Radio Paradise diversity examples
High Diversity Playlists

Artist Track Tags
Big Head Todd &
It’s Alright rock, alternative, jam band, prog rock, Jam, 90s
The Monsters

folk, singer-songwriter, female vocalists, Canadian,
Joni Mitchell Be Cool classic rock, acoustic

jazz, trumpet, cool jazz, blues, jazz vocals, easy
Chet Baker Tangerine listening

Diversity: 1.0
18 unique tags out of 18
170

Pandora diversity examples

Artist Track Tags
Project industrial, ebm, electronic, darkwave, Gothic,
Timekiller
Pitchfork synthpop,

melodic black metal, black metal, synthpop, metal,
Covenant We stand alone industrial, futurepop

ebm, industrial, futurepop, electronic, synthpop,
Icon of Coil Faith? Not Important darkwave

ebm, futurepop, industrial, synthpop, electronic,
Neuroticﬁsh Waving Hands goth

Project industrial, ebm, electronic, darkwave, Gothic,
Momentum
Pitchfork synthpop

melodic black metal, black metal, synthpop, metal,
Covenant Stalker industrial, futurepop

Diversity: 0.305 Project Pitchfork Radio

Pandora diversity examples

High Diversity Playlists

Artist Track Tags

metal, thrash metal, heavy metal, rock, hard rock,,
Metallica The Call of Ktulu metallica

Linkin Park Pushing Me Away rock, Nu Metal, alternative, metal, Linkin Park, punk

Creed One Last Breath rock, alternative, hard rock, Grunge, metal, punk

Diversity: 0.611 Evanescence Radio

Musicmobs diversity examples

Artist Track Tags
rock, alternative, Progressive rock, metal, hard
Perfect Circle (54 Tracks) rock, industrial

Progressive metal, Progressive rock, metal, rock,
Tool (43 Tracks) alternative, Progressive

Diversity: 0.014
8 unique tags out of 582
173

Playlist Cohesion Metric

Z
1
• Goal - ﬁnd level of cohesion
Y 3
R
Q in an ordered sequence such
2 5
S
4
5 as a playlist
2
W X

•
4 4 P
4 T O
2
How:
3

V
2 U

J
5 N 1

M • Represent the item space
3
I
3 3 5 as a connected graph
1 K

H 1 5
G 2
L
• Find the shortest weighted
3 E
2
2 path that connects the
F

A B
1
4 1
ordered sequence
2
3
C 1
D
• Average step length is the
cohesion index

174


Z
1 R
Y 3 Q
2 5
S
5
4
2
W X 4 4 P

•
4 T O

2
3
5
2

1
Consider [A, E, U, X]
U N

•
V M
3
3 J
3 5 Distance: [3,7,6] = 16
I
K

•
1
H 1 5
G 2
L
Average Distance: 5.33
2
3 2
E
F
1
A B 4 1
2
3 D
C 1

175


Z
1 R
Y 3 Q
2 5
S
5
4
2
W X 4 4 P

•
4 T O

2
3
5
2

1
Consider [Z,L, H, X]
U N

•
V M
3
3 J
3 5 Distance: [15 , 10 , 9] = 34
I
K

•
1
H 1 5
G 2
L
Average Distance: 11.3
2
3 2
E
F
1
A B 4 1
2
3 D
C 1

176

Building the graph
MusicBrainz Artist Relations
• Nodes are artists

• Edges are relations, weighted by signiﬁcance

• 132 Relationship types. some examples:

Edge type Weight
Is Person 1
Member of band 10
Married 20
Performed with 100
Composed 250
Remixed 500
Edited Liner Notes 1000
177

MusicBrainz Artist Relations Graph
Average inter-song
Source
Distance
Radio Paradise 0.08 / 0.06
Pandora 0.11 / 0.12

MusicMobs 0.13 / 0.10

Art of the mix 0.14 / 0.10

Random (RP) 0.27 / 0.22

Random (graph) 0.39 / 0.45

Random (AotM) 0.56 / 0.19

178

Building the graph
Echo Nest Artist Similarity

• Nodes are artists

• Edges are similar artists, weighted by similarity

179

Echo Nest Artist Similarity Graph
Average inter-song
Source
Distance
Pandora 1.57 / 1.4
Radio Paradise 2.27 / 1.0

MusicMobs 2.71 / 1.7

Art of the mix 3.02 / 1.4

Random (RP) 4.02 / 1.2

Random (AotM) 7.00 / 1.1

Random (graph) 7.89 / 1.78

180

Hybrid Radio
The Social Radio
• produce playlists via weighted distance
paths
• next destination song is determined via a
vote across all listeners
• candidate songs selected from disparate
communities

Hybrid Radio
Ratings

• ratings are applied to the edge that lead to
the song
• song ratings -> playlist ratings
• serving 2 purposes
• direct evaluation of playlists
• object based ﬁltering
184

Convergence

When the cloud provide all the music and
ubiquitous internet provides it all the time
recommendation and playlisting merge

186

Convergence

The celestial jukebox needs a DJ.

187

The anonymous programmers who write the
algorithms that control the series of songs in these
streaming services may end up having a huge
effect on the way that people think of musical
narrative—what follows what, and who sounds best
with whom. Sometimes we will be the d.j.s, and
sometimes the machines will be, and we may be
surprised by which we prefer

You, the D.J.
Online music moves to the cloud.
by Sasha Frere-Jones
The New Yorker, June 14, 2010 188

Finding a Path Through the Juke Box: The Playlist Tutorial

More Related Content

What's hot (11)

Viewers also liked (8)

Similar to Finding a Path Through the Juke Box: The Playlist Tutorial (20)

Recently uploaded (20)

Finding a Path Through the Juke Box: The Playlist Tutorial