VISUALIZING NETWORKS OF MUSIC ARTISTS WITH RAMA

Lu´ıs Sarmento

, Fabien Gouyon

, Bruno G. Costa

and Eug´enio Oliveira

LIACC/FEUP, Univ. do Porto, Rua Dr. Roberto Frias, s/n, Porto, Portugal

INESC Porto, Rua Dr. Roberto Frias, 378, Porto, Portugal

Univ. Cat´olica Portuguesa, Rua Diogo Botelho, 1327, Porto, Portugal

LIACC/FEUP, Univ. do Porto, Rua Dr. Roberto Frias, s/n, Porto, Portugal

Keywords:

Information visualization, Search interfaces, Content ranking using social media, User interfaces for search

interaction.

Abstract:

In this paper we present RAMA (Relational Artist MAps), a simple yet efﬁcient interface to navigate through

networks of music artists. RAMA is built upon a dataset of artist similarity and user-deﬁned tags regarding

583.000 artists gathered from Last.fm. This third-party, publicly available, data about artists similarity and

artists tags is used to produce a visualization of artists relations. RAMA provides two simultaneous layers of

information: (i) a graph built from artist similarity data, and (ii) overlaid labels containing user-deﬁned tags.

Differing from existing artist network visualization tools, the proposed prototype emphasizes commonalities

as well as main differences between artist categorizations derived from user-deﬁned tags, hence providing

enhanced browsing experiences to users.

1 INTRODUCTION

One of the fastest growing media on the web is

web-radio. There are now many web-radios avail-

able where millions of users spend a very signiﬁ-

cant amount of their time. Users can customize a

million-track collection to very speciﬁc music tastes.

Web-radios usually allow users to type in a tag that

describes the music they want to hear (e.g. “acid

jazz”, “wake up”, etc.), and music items with that

tag will make up the personalized radio feed recom-

mended to that user. Artist similarities (computed

by methods described in Section 2) are also used

to generate playlists. Albeit very useful, tag-based

or similarity-based playlists are sometimes difﬁcult

for users to understand. The reason why a given

artist or music was “selected” by the web radio is

not always obvious and it can easily become confus-

ing or frustrating for less experienced users that are

unable to clearly express their musical preferences

through queries. For example, very famous artists

(e.g. “U2”) can sometimes be considered “similar” to

other –otherwise quite different– popular artists (e.g.

“Queen”, “Sting”, “Coldplay”, “Counting Crows”),

just because they are also popular or end up receiving

the same relatively uninformative tags (e.g. “pop”).

In this paper, we explore the idea that net-

works of music artists contain rich and multi-

faceted information (music artist similarities, user

tags, etc.) that can be useful for recommendations

that go beyond the creation of playlists. We present

RAMA, Relational Artist MAps, available through

http://pattie.fe.up.pt/RAMA/. RAMA is a visualiza-

tion tool that allows the user to navigate inside the

network of artists of the Last.fm web-radio. Our tool

uses information about artists similarity and artists

tags provided by Last.fm to produce a visualization of

artists relations and corresponding user-deﬁned tags

in a graph, hence fostering the visualization of rela-

tional information. One of the original contributions

of our work over Last.fm’s web interface is to allow

users not only to see which tags are common to a set

of artists, but also those which are speciﬁc to a given

artist when compared to similar artists.

2 RELATED WORK

There are different approaches to music recommen-

dation, which makr use of the different facets of mu-

sic. Recommending music items is essentially based

on the deﬁnition of a similarity metric between items,

possibly tuned for a particular user. As developed in

greater details in (Celma, 2006), this can be achieved

232

Sarmento L., Gouyon F., Costa B. and Oliveira E.

VISUALIZING NETWORKS OF MUSIC ARTISTS WITH RAMA.

DOI: 10.5220/0001820802320237

In Proceedings of the Fifth International Conference on Web Information Systems and Technologies (WEBIST 2009), page

ISBN: 978-989-8111-81-4

by: (i) automatic analysis of contents (i.e. via algo-

rithms computing e.g. rhythmic and tonal descrip-

tions of audio ﬁles), (ii) expert analysis (as is the

case of Pandora, or Tapestry

), (iii) collaborative ﬁl-

tering (i.e. exploiting a user-item relational matrix),

(iv) co-occurrence analysis (using e.g. crawling tech-

niques to fetch from the web text related to mu-

sic artists and seek co-occurrences of artists names,

terms, etc. (Schedl et al., 2007)), (v) analysis of

meta-information provided by users (i.e. recommend-

ing music clips with a speciﬁc tag –e.g. “alternative

rock”). Hybrid alternatives to these techniques also

exist.

The most frequent way for systems to present mu-

sic recommendations to users (e.g. on the Last.fm

website), is by generating lists of potentially relevant

music items (artist names, songs, etc.). An alterna-

tive way, is to provide users with more informative

visualization methods for inspecting the similarities

between music items (Pampalk et al., 2002), (Pam-

palk and Goto, 2007). For instance, a lot of attention

has recently been given to the visualization of artist

networks. In artist networks, two different types of

data are usually presented. On the one hand, there is

traditional encyclopedic data about individual artists.

This includes biographic data such as names of al-

bums, names of music tracks, photos and other im-

ages, etc. On the other hand, there is relational data,

regarding e.g. artist similarities, connections between

them, etc.

A popular metaphor for visualizing these two

types of data is that of connected graphs, where the

data is presented through nodes and edges connect-

ing them. Connected graphs offer a number of “con-

tainers” to represent information (Shneiderman and

Aris, 2006): (i) Node Labels, (ii) Node Attributes,

(iii) Edge Labels, (iv) Edge Attributes (v) Edge Di-

rections (either directed or undirected links). Given

speciﬁc types of artist-related data one wants to visu-

alize, a speciﬁc mapping must be done onto these data

“containers”. Nodes (or vertices) and edges are also

central to the science of complex networks (Barab´asi,

2002), and a number of works in this ﬁeld have re-

cently brought some light onto the manifold inter-

twinements of musical artists networks (Cano et al.,

2006), (Teitelbaum et al., 2008). There are several

applications for visualizing artist networks as two-

dimensional connected graphs. For instance Musi-

covery

, TuneGlue’s “music map”

, Gnod’s “music

http://www.amgtapestry.com/

http://musicovery.com/

http://audiomap.tuneglue.net/

map”

, Dimvision’s “music map”

, Kyle Scholz’s

music recommendation tool

, or the “SimilarArtist-

Graph”

by Last.fm user Shoxrocks. These applica-

tions make use of both individual and relational data.

In some applications, nodes can be expanded to reveal

attributes such as label names, dates of album release,

biographies and link to artist websites.

In the above-mentioned applications, data regard-

ing the similarity between artists is gathered from

third parties such as Last.fm or Amazon. Artists (i.e.

nodes) that are somehow similar (i.e. that Last.fm

considers similar, or that Amazon recommend to-

gether) share a link (i.e. are connected via an edge).

A bird eye’s view upon the topology of artist

graphs may reveal clusters of artists, and users can

embrace in one sight artists that are similar to their

query as well as those similar to the answers, and so

on.

Our system also uses third-party (i.e. Last.fm)

similarity data to display music artists in 2D con-

nected graphs, where edges represent similarities. In

our prototype, node labels also convey similar data

as they do. But an original aspect is that we intend to

make further use of node attributes, as well as edge la-

bels, edge attributes and directions, in order to convey

more relational information (i.e. artist commonalities

and speciﬁcities) computed from Last.fm data. By

making use of more data, we sought a good balance

between readability (avoiding cluttered use of space)

and richness of the data presented to the user. Hence

the special focus, in the design phase, on a proper use

of graphical features (e.g. colors and transparencies)

as well as interactivity between the user and the pro-

totype (some information is shown by default, some

other only as results of users’ interactions).

3 Last.FM

Last.fm (http://www.last.fm) is one of the leading

internet-based social music platforms, where users

can listen to music, ﬁnd information about artists they

like, or discover artists they might not know. Follow-

ing Web 2.0 concepts, users can also set up their own

proﬁle, facilitating targeted automatic recommenda-

tions, among other things they can also get informa-

tion about users with similar tastes, gigs in their lo-

cal area, videos, etc. Last.fm provides an interface for

users to collaborativelyedit encyclopedic information

http://www.music-map.com/

http://www.dimvision.com/musicmap/

http://kylescholz.com/projects/speaking/tae2006/music/

http://lastfm.dontdrinkandroot.net/

VISUALIZING NETWORKS OF MUSIC ARTISTS WITH RAMA

233

about artists. For some of the more popular artists

there is also extensive biographic information already

available.

User proﬁles –hence recommendations– are con-

stantly updated via a software (free of use) which

gathers (“scrobbles” in the Last.fm vernacular) statis-

tics about the music listened to by users. User listen-

ing patterns are recorded and analyzed by Last.fm in

order to better organize and recommend music. Users

are also encouraged to organize the music they lis-

ten to by assigning tags to artists, or even to spe-

ciﬁc albums or tracks. The deﬁnition of tags is up

to the users and can describe any aspect users believe

are relevant, as music genres (e.g. “rock”, “Viking

metal”), locations (e.g. “Berlin”), mood (e.g. “chill”),

opinions (e.g. “songs my mother would like”), con-

texts (e.g. “love”) or just about anything that cross

users’ minds (see Figure 1 for examples of tags as-

signed to the band “Radiohead”). Tagging music

helps users to browse Last.fm contents. But the real

power of tags becomes clear when considering that

tags of hundreds of thousands of users are combined,

providing an emerging “bottom-up” categorization of

music.

Figure 1: List of user tags for the band “Radiohead”. Sizes

of the letters correspond to popularity of the tag.

A cornerstone of Last.fm functionalities resides in

links of similarity between artists (which can be seen

on Figure 2 (a simple list), and which is central to au-

tomatic recommendations made to users). The algo-

rithm used for computing similarities between artists

is unknown –to the authors of this paper– but is prob-

ably based on (i) the analysis of user-added tags, (ii)

user listening patterns such as co-occurrence statistics

(users that listen to artist X also listen artist Y), and

(iii) user proﬁles information (“similar” users should

like “similar” artists).

4 RAMA: RELATIONAL ARTIST

MAPS

RAMA is built on top of a client-server architecture.

The visualization is performed on the client side (the

user application) using information obtained from the

Figure 2: Personal artist recommendations made by Last.fm

to one of the authors of this paper.

server via an HTTP request. The server manages the

data that has been extracted from Last.fm site and per-

forms all the necessary pre-processing operations to

provide the client with the information needed for vi-

sualization.

Figure 3: System overview.

Given an initial query (i.e. an artist name) submit-

ted by the user through a text-box in the client appli-

cation, a request is sent to the server which provides

all the information that the client application requires

to draw the corresponding artist network. The net-

work contains the artists that are found to be more

similar to the queried artists, according to Last.fm’s

artist similarity index. It also contains the artists that

are found by propagating the similarity query to the

previously found ones so that we obtain artists sev-

eral levels away from the initial artist. This allows

the user to see the artist position in wider context (see

e.g. Figure 4). The server provides the client with

a list containing information regarding each node in

the graph (i.e. each artist in the network), namely: (i)

the 2D coordinates of the corresponding node in the

graph, computed by our graph layering algorithm (see

Section 4.2); (ii) the list of user-deﬁned tags for each

artist as taken from Last.fm site; (iii) explicit similar-

ity relationships with other artists in the network (for

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

234

drawing the necessary edges) as taken from Last.fm

site. Once the visualization is rendered by the client

application, users can either insert another query in

the text-box, or browse through the network, inter-

actively consulting tag information, and performing

additional queries by clicking on the nodes. Interac-

tion continues in this fashion until the user quits the

application.

4.1 Data Pre-processing

Our system uses data that is freely available through

Last.fm web-services API

. Through this web-service

developers can access different data categories related

to Last.fm users and media. Available data includes:

(i) the proﬁle of users of Last.fm, (ii) information

and statistics related to artists, their albums and corre-

sponding music tracks, (iii) tag information for each

of the previous items, (iv) information about user cre-

ated groups (e.g. fan groups), (v) information about

message forums that users can create/participate, and

(vi) geo-aware statistics about users and music pref-

erences (more details can be found in the web service

site). From all the data available, our system uses cur-

rently only data specially concerning artists, namely:

• basic artist data: name, URL of image, “popular-

ity” index within Last.fm community,

• the list of the most similar artists for each artist.

Last.fm assigns each similar artist a weight rang-

ing from 100 (full similarity) to 0 (almost no sim-

ilarity).

• information about the user-deﬁned tags for each

artist. Again, Last.fm assigns weights to quantify

the association level for that tag: 100 means full

association, while 0 indicates loose association.

Because each individualaccess to the Last.fm web

service involves a considerable overhead due to net-

work latency and server load, we created a local

copy of the data we needed by crawling the web-

service systematically. For a period of about a week

(from January 30 to February 4 2008) we consulted

Last.fm’s web-service and obtained the previously de-

scribed data for a total number of 583.000 artists. Lo-

cal access to this data allowed to speed-up experi-

mentation and to improve the global performance of

our system. Although the Last.fm user community

is constantly contributing and changing this informa-

tion, we can assume that the relationships between

artists, which are the focus of our work, are more or

less stable, at least for the reasonably popular artists.

In any case, the crawling procedure can be repeated at

any time in order to update data.

http://www.audioscrobbler.net

4.2 Layering the Graph

For layering the graph corresponding network of

artists in a 2D plane, we implemented our own graph

drawing system based on a force-directed placement

strategy (Fruchterman and Reingold, 1991). Our im-

plementation, in Perl, is completely integrated in the

overall server-side framework which was also imple-

mented in Perl. There are several parameters related

to the graph layering operation that need to be set in

order to create an appealing and useful visualization.

The ﬁrst parameter concerns the number of nodes (i.e.

artists) that should be included in the graph. We wish

to convey as much information as possible but we are

limited to a given frame size. Layering graphs with

many labeled nodes –which may have other attributes

to be drawn– in a regular computer screen may gen-

erate rather confusing visualizations. Therefore we

need to limit the number of nodes to a reasonably low

value, for example between 15-50 nodes.

The second parameter has to do with the size of

the context we wish to convey. Since we are limit-

ing the number of nodes to a given maximum, there is

a balance between (i) choosing more nodes that are

closer to the initial artist, and thus providing more

detail about the close relations of that artist, or (ii)

choosing more nodes that are a few links away from

the artist and thus provide a wider perspective about

the artist and its relations.

Our goal is to combine a good level of local detail

with a wide enough perspective, and achieve a good

balance between readability and richness of data pre-

sented to the user. We have deﬁned 3 parameters to

conﬁgure the properties of the network to be drawn

around a given artist:

• Level-1 BranchingFactor: this parameter controls

how many nodes directly connected to the initial

one are to be visualized (the nodes chosen are the

top similar ones).

• Level-n BranchingFactor: this parameter controls

how every node in the graph except the initial one

will branch.

• Maximum Branching Distance: this parameter

imposes a threshold on the maximum branching

distance, i.e. how many links away can nodes be

from the initial artist node.

With these parameters we perform an iterative expan-

sion of the initial artist node, branching each node ac-

cording to the corresponding factor. Increasing level-

1 branching will promote local detail, while increas-

ing level-n-branching and maximum branching dis-

tance will widen the context. In any case the number

VISUALIZING NETWORKS OF MUSIC ARTISTS WITH RAMA

235

of nodes can never exceed a given pre-deﬁned thresh-

old.

4.3 User Interface

The user interface was developed in Processing

and

is responsible for generating the visualization and

providing interactivity. Initially, the user can enter a

name of an artist in a text-box. A query is then sent

via HTTP to the server, which replies with the artist

network data. For each artist (i.e. each node) in the

network, the server sends back to the client the fol-

lowing data: (i) the name, (ii) a popularity index, (iii)

the url of the photo of the artist in Last.fm server, (iv)

the top 20 user-deﬁned tags, (v) information about

similarity with other artists (i.e. the node edges and

their weights), and (vi) the coordinates for placing the

corresponding node of the graph in a 2D plane. All

this information is sent to the visualization interface

in text format to allow a simple parsing procedure.

The interface application uses additional data that

is directly fetched from Last.fm site at runtime,

namely artist pictures and a short biographic descrip-

tion. For these media items performance constraints

are not so severe (only two extra accesses for each

artist in the network), so the interface can access that

information directly as needed without requiring any

intervention from the server.

On top of the basic network structure we place in-

formation related to user-deﬁned tags. However, in-

stead of presenting all the tags assigned to each artist

overits correspondingnode, we try and place the most

common tags in the network in a such position that

they will simultaneously describe all artists contained

in a certain part of the network (simple animations of

these tags permit to avoid readability problems). Such

tag information is shown by default, no action be-

ing required from the user, allowing the user to easily

identify the attributes that explain why certain artists

are clustered in a speciﬁc part of the network.

For instance, in the network of “Radiohead” (see

Figure 4), one can identify a region of artists tagged

as “alternative” in the center because such a tag is

common to most artists in the whole network. On the

other hand, the region on the top-left branch is tagged

“electronica”, the common attribute of artists in that

part of the network. Tag sizes are proportional to the

number of artists for which they are relevant.

Moving the mouse cursor on top of a speciﬁc artist

node results in the presentation of diverse data: the

artist picture, a short description and bio (gathered

at run-time from the Last.fm site), and a link to its

Last.fm webpage. Left-clicking on any artist name

http://www.processing.org

Figure 4: Example graph for artist “Radiohead”.

has the same effect as entering its name in the query

box, i.e. sending a new query and refreshing the map

with this query as seed. This allows simple user nav-

igation through the artist network. To ensure that vi-

sualization always exhibits some degree of novelty, a

new graph layout is computed in real time for each

query, even if the query has been processed before.

Figure 5: Examples of speciﬁc tags for artists “Bjork” and

“Massive Attack”, both part of the “Radiohead” network.

An original feature of our user interface resides in

the possibility to visualize tags that are speciﬁc to an

artist in relation to another, i.e. those tags that are

relevant to an artist but not to its neighbors. This in-

formation is shown only when the user crosses the

mouse cursor over an edge. This action will result

in the rendering of the tags that are speciﬁc to each

of the two artists at their respective sides of the edge.

In Figure 5 one can see an example of such a behav-

ior. “Bjork” and “Massive Attack” are both part of the

“Radiohead” network. Although they are both on a

region characterized by the “electronic” tag, the tags

“icelandic”,“female” and “avant-garde” are only as-

signed to “Bjork”, while the tags “UK”, “british” and

WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies

236

“chill” are assigned only to “Massive Attack”. This

provides a very clear description of the unique fea-

tures of each artist in the network. We believe this

to be a very powerful tool in increasing user under-

standing about the network data. Table 3 shows a few

more examples of tags that are found to be speciﬁc to

a given artist (“Radiohead”) when compared to other

artists in its network.

Table 1: Pairwise discriminative tags between “Radiohead”

and other similar artists.

Radiohead Sigur Ros

UK, pop, britpop, art

rock

shoegaze, chillout, ice-

landic, ethereal

Radiohead Thom Yorke

progressive rock, 90s,

post-rock, experimental

rock

radiohead, trip-hop,

electro, chillout,

singer-songwriter, male

vocalists

Radiohead Placebo

art rock, experimental,

emo, glam rock, metal,

punk

Radiohead Muse

electronica, art rock,

post-rock

progressive, emo, metal

5 CONCLUSIONS AND FUTURE

WORK

RAMA provides a simple yet efﬁcient interface to

navigate through the network of similar artists, allow-

ing users to obtain a wider view about the artists they

know, and toeasily discover newbands and artists that

they might like. It provides two simultaneous layers

of information: (i) a graph built from artists and their

connections, and (ii) overlaid labels containing user-

deﬁned tags that express the classiﬁcation made by

Last.fm community for each of the artists. From ex-

perimentation we have observed that the system effec-

tively allows to identify clusters of tightly connected

band and artists (such as for example former mem-

bers of a band that pursued a solo career). Addition-

ally, the visualization procedure in RAMA also em-

phasizes the main differences between artists, allow-

ing the user also to visualize which are the most dis-

tinctive attributes of each artist.

Future work includes enhancing user experience

by adding song snippets for each artist, so that the user

can play them on demand while navigating across the

network. Also, we plan to improve interactivity by

allowing the user to optionally navigate though user-

deﬁned tags, and not just artists. We will also focus

on allowing the user to manipulate the graph (zoom-

ing, rotating, etc) and to edit it. Editing capabilities

will enable the user to remove nodes (artists) from the

graph, expand only some, and thus generate a person-

alized graph, which could then be saved e.g. in the

form of a playlist.

ACKNOWLEDGEMENTS

This work was partially supported by grant

SFRH/BD/ 23590/ 2005 from FCT (Portugal),

co-ﬁnanced by POSI. Thanks to Elias Pampalk from

Last.fm for insightful comments.

REFERENCES

Barab´asi, A. (2002). Linked: The New Science of Networks.

Perseus Publishing.

Cano, P., Celma, O., Koppenberger, M., and Buld´u, J.

(2006). The topology of music recommendation net-

works. Chaos, 16(013107).

Celma, O. (2006). Music recommendation: A multi-faceted

approach. Master’s thesis, Universitat Pompeu Fabra,

Barcelona.

Fruchterman, T. M. J. and Reingold, E. M. (1991). Graph

drawing by force-directed placement. Softw. Pract.

Exper., 21(11):1129–1164.

Pampalk, E. and Goto, M. (2007). MusicSun: A new ap-

proach to artist recommendation. In Proceedings of

the International Conference on Music Information

Retrieval.

Pampalk, E., Rauber, A., and Merkl, D. (2002). Content-

based Organization and Visualization of Music

Archives. In Proceedings of the ACM Multimedia,

pages 570–579, Juan les Pins, France. ACM.

Schedl, M., Knees, P., Widmer, G., Seyerlehner, K., and

Pohle, T. (2007). Browsing the Web Using Stacked

Three-Dimensional Sunbursts to Visualize Term Co-

Occurrences and Multimedia Content. In Proceedings

of the 18th IEEE Visualization Conference.

Shneiderman, B. and Aris, A. (2006). Network visualiza-

tion by semantic substrates. IEEE Transactions on Vi-

sualization and Computer Graphics, 12(5).

Teitelbaum, T., Balenzuela, P., Cano, P., and Buld´u, J.

(2008). Community structures and role detection in

music networks. Chaos, 18(043105).

VISUALIZING NETWORKS OF MUSIC ARTISTS WITH RAMA

237