[ieee 2009 international conference on advances in social network analysis and mining (asonam) -...
TRANSCRIPT
Mining the Dynamics of Music Preferences from aSocial Networking Site
Nico Schlitter
Faculty of Computer Science
Otto-v.-Guericke-University Magdeburg
Magdeburg, Germany
Tanja Falkowski
Faculty of Computer Science
Otto-v.-Guericke-University Magdeburg
Magdeburg, Germany
Abstract—In this paper we present an application of ourincremental graph clustering algorithm (DENGRAPH) on a dataset obtained from the music community site Last.fm. The aim ofour study is to determine the music preferences of people and toobserve how the taste in music changes over time. Over a periodof 130 weeks, we extract for each interval user profiles of 1,800users that represent their music listening behavior. By buildingand incrementally clustering a graph of similar users, we obtaingroups of people with similar music preferences. We label theseclusters with genres according to the user profiles of the clustermembers. Due to the incremental nature of DENGRAPH we showhow clusters evolve over time. Besides the growth and decrease ofclusters we observe how new clusters emerge and old clusters die.Furthermore, we show the merge and split of clusters. The resultsof our experiments indicate that DENGRAPH is particularlyuseful to efficiently detect groups of similar users and to trackthem over time.
Keywords-Social Networks; Network Dynamics; CommunityAnalysis; Music Preferences
I. INTRODUCTION
Music is a constant companion for most people. In 2000,
Eurostat, the statistical office of the European Communities,
carried out a survey on Europeans’ participation in cultural
activities, one part of which concerned music. The results
show, for example, that over 60% of Europeans listen to musicevery day [1]. Music is known to induce emotions such as
’happiness’, ’sadness’ and ’fear’ [2] and studies indicate that
the meaning of music in the lives of people changes over time.
Younger people tend to link music preferences to a variety
of psychological characteristics such as personality, personal
qualities and values. Since, individuals have clearly defined
stereotypes about the fans of various music genres, music is
used to convey information about oneself to observers [3]. On
the other hand, for older people, music often contributes to
positive aging by providing ways to maintain positive self-
esteem, feel independent and avoid feelings of isolation or
loneliness [4]. Based on this observation, the hypothesis is
that the music preferences of people change as well over their
lifetime.
Changes in music preferences are often triggered by the
environment a person lives in, such as media or friends. Many
people appreciate of being introduced to works that they do not
know of yet. This is a reason why recommendation systems for
music that use e.g. collaborative filtering techniques recentlybecame more and more popular (see, e.g., [5]). Collaborative
filtering is a procedure that detects the preferences of users
for items such as books, DVDs or music. These preferences
are often represented in user profiles. The profiles can becollected either explicitly or implicitly. In the first case, a user
is explicitly asked to rate items that he has purchased or that
he likes. An implicit profile is based on observations that have
been made based on interaction data (such as purchases). In
this study we build for each user in different points in time,
implicit profiles which represent the current music listening
preferences.
In this paper, we apply DENGRAPH, a clustering algorithm
we presented in [6], to study the temporal dynamics of the
music preferences of 1,800 individuals. We use data obtained
from the social networking site Last.fm, which provides listsof the most listened artists for each week over the lifetime
of a user. Based on this information we build user profiles
by extracting the genres of the most listened artists. The
artist’s genre is determined by the tags that the community
members use to characterize the artist. The profile of a user
is then represented by a vector of the weighted tags of his
favorite artists. We represent users as nodes in a graph and
connect them with an edge, if their profile similarity reaches
a predefined threshold. Due to the incremental nature of
DENGRAPH, temporal changes invoke an clustering update.
By this we are able to observe and analyze when and how the
taste in music of an individual user changes over time and how
his membership to a certain community changes accordingly.
The results of our study indicate that DENGRAPH is capable
to efficiently extract communities of users with a similar taste
in music and that it furthermore allows us to monitor and
analyze the changes in the listening behavior of individuals and
the resulting shifts in the membership to a certain community.
The rest of the paper is organized as follows. In Section II
we present our approach to represent the music preferences
of users in user profiles. Section III discusses how clusters
of similar users can be detected and observed over time
using the density-based clustering algorithm DENGRAPH.
The characteristics of the used data set and the results of
our experiments are presented in Section IV. We conclude
the paper in Section V.
2009 Advances in Social Network Analysis and Mining
978-0-7695-3689-7/09 $25.00 © 2009 IEEE
DOI 10.1109/ASONAM.2009.26
243
2009 Advances in Social Network Analysis and Mining
978-0-7695-3689-7/09 $25.00 © 2009 IEEE
DOI 10.1109/ASONAM.2009.26
243
II. REPRESENTING MUSIC PREFERENCES
Last.fm1 is a music community with over 20 million activeusers based in more than 200 countries. After a user signs
up, a plugin is installed and all tracks a user listens to are
submitted to a database. To study the temporal evolution
of music listening behaviors, we use this information and
represent a user’s taste in music in form of user profiles. This
process is described in the following section. Furthermore, to
find groups of similar users, we discuss how a graph of similar
users is constructed.
A. Building User Profiles
In affiliation networks, users are connected based on a
certain property that they share. This property can be for
example a similar taste in music. A community is then defined
as a group of people who share similar music preferences. We
define a user by a profile that describes his music preferences
and determine communities of users by clustering similar
profiles.
Genre labels are often used as ground-truth for describing
song or artist similarity in music information retrieval (MIR)
systems. A genre is a rather high level label for an artist
and depends on a variety of aspects such as the rhythm,
melody or used instruments. Genres are usually predefined by
producers or the artist and the resulting classes are highly over-
lapping. Developing automatic genre classification systems is
a controversial (see, e.g., [7]) but still popular topic in music
information retrieval and remains a challenging task (cf., e.g,
[8], [9], [10]).
In contrast to using a single label, McKay and Fujinaga
[11] suggest to use a ranked list of multiple genres to label
one recording. Similar to this suggestions, Geleijnse et al.
[12] propose to use tags that result from community Websites
where users collectively attribute artists with labels. Contrary
to expert-defined genre classification, a community-based tag-
ging results in a ranked list of terms that is richer than a
single genre. Experiments have shown that the Last.fm tagsare descriptive and consistent with respect to their similarity
and that they present a valuable characterization of artists and
their music [12].
For each artist, Last.fm provides the tags that are used byusers to describe the artist. For example, the band ’ABBA’
has 41 tags. The most often used tags are ’pop’, ’disco’,
’swedish’ and ’70s’. The singer ’Amy Winehouse’ has 34 tags,
the most often used tags are ’soul’, ’jazz’, ’female vocalists’
and ’british’. We use the most often assigned genre tags for
artists, merge tags that are semantically identical (such as ’hip
hop’ and ’hip-hop’) and obtain a set G of 222 genre tags gi.The provided tags gi are used to describe the artists in a
genre vector as follows:
Definition 1: A = {�a1, . . . ,�an} is a set of artists. An artistai is defined as a genre vector �ai = (g1, g2, . . . , gk) of thek-most used genre tags gi ∈ G = {g1, g2, . . . , gk}, where Gis the set of genres.
1http://www.last.fm/
For each user p we obtain from Last.fm for each interval ta list of the most listened to artists including the number of
times the respective artist was heard by the user (playcount).The user profile in interval t is defined as follows:
Definition 2: For each user p in interval t, a user profile isdefined as
�pt =m∑
i=1
�ai · playcounti, (1)
where m is the number of artists listened to in the interval t.Thus, each genre element is weighted by the number of times
played. The weighted vectors are added and the profile is a
vector representing the user’s music preferences.
B. Building a Graph of Similar Users
To find groups of users with similar music preferences, we
first build a similarity graph. An edge in the similarity graph
indicates that two users are similar to a certain extend. We
determine the similarity of two profiles by using the cosinesimilarity. The cosine of the angle θ between two vectors a, bcan be interpreted as a measure of similarity and is calculated
as follows: cosθ = (a · b)/(|a||b|). A cosine value of zeroindicates that the two profile vectors are orthogonal, a value
of one indicates an exact match. The similarity between the
user p1 and p2 is then defined as follows:Definition 3: The similarity between two users p1 and p2
is defined as
sim(p1, p2) =∑m
i=1 p1i · p2i√∑mi=1(p1i)2 ·
∑mi=1(p2i)2
(2)
Using this formula, we calculate the pairwise similarity of
user profiles for each interval t. Since we are only interestedin profiles that are similar to a certain extend, we establish
links between nodes only if a similarity threshold δ is reached.Finally, we obtain an unweighted and undirected graph of
similar users.
Definition 4: The similarity graph G is represented by a listof nodes V = {p1, . . . , pn} and a list of edges E = {(px, py)},while sim(px, py) ≥ δ, ∀(px, py) ∈ E.
III. DETECTING COMMUNITIES AND EVOLUTION
To find and track communities of similar users with respect
to their music preferences, we apply our incremental DEN-
GRAPH clustering algorithm presented in [6]. We extend the
algorithm to allow for individuals being a member of more
than one cluster. In the following section, we briefly describe
the clustering algorithm and the incremental update procedure
to observe clusters over time. Furthermore, we present how
the cluster labels are determined.
A. Incremental Graph Clustering
The intention of DENGRAPH is to cluster similar nodes in
a graph into communities. The density-based approach applies
a local cluster criterion. Clusters are regarded as regions in the
graph in which the nodes are dense, and which are separated
from regions of low node density. To detect regions of higher
density, DENGRAPH computes neighborhoods which have a
244244
directly density reachable
p is
p
q
from q
density reachable
p is
pq
from q
density connected
p is
from q
core /border
core
p
q
m
p (q) and (q)|
Fig. 1. The concepts directly density reachability, density reachability anddensity connectedness to determine whether nodes are density connected.
given radius (ε) and must contain a minimum number of nodes(η) to ensure that the neighborhood is dense. A node that hassuch a neighborhood is termed a core node. Nodes that haveno such neighborhood are either border nodes if they are inthe neighborhood of a core node or noise nodes.To build a cluster, DENGRAPH traverses the graph by
randomly picking nodes and places all density connected (cf.Figure 1) nodes it encounters to the same cluster. If a node is
not density connected to the nodes seen thus far, it is assigned
to the next cluster candidate. Not each node becomes member
of a cluster: If a node does not have an adequately dense
neighborhood w.r.t. ε and η and is not density connected toany other node, then it is termed a noise node and its clustercandidate is dropped.
In Figure 1, the concepts of direct density reachability,
density reachability and direct connectivity are illustrated. For
a more detailed description cf. [13], [14], [6].
To allow for tracking and analyzing the temporal dynam-
ics of clusters, DENGRAPH is designed as an incremental
procedure: The clustering is updated incrementally based on
the changes that are observed in the graph structure from one
interval to another. Changes in the graph structure may evoke
one of the following clustering updates: (1) creation of a newcluster, (2) removal of a cluster, (3) absorption of a new clustermember, (4) reduction of a cluster member, (5) merge of twoor mores clusters and (6) split of a cluster into two or moreclusters. For a more detailed description of the incremental
graph clustering algorithm cf. [6].
A potential split of a cluster is handled before the update
method shown in Figure 2 is invoked if an edge between two
core nodes has been removed.
If no split has been observed, we check for each node
that has changed, whether the respective node p has an ε-neighborhood with more than η neighbors (Np(ε) ≥ η). Ifis has a neighborhood, we check whether (i) a new cluster is
established, (ii) the node becomes a member of another cluster
or if (iii) two or more clusters merge. If the node has no ε-neighborhood, we examine whether (i) a cluster is removed,
(ii) a cluster member has been removed, or (iii) a node has
been absorbed by another cluster. An overview of the updatefunctions is shown in Figure 2.
number of cores in N( )with distinct clusterID
Np( )
Creation MergeAbsorption
number of cores in N( )with distinct clusterID
p is core node p is core node
Removal Reduction Reduction Removal / Reduction / Absorption
n0
falsetrue
0 n1
falsetrue falsetrue
Fig. 2. The update Method to detect cluster transitions. (A potential split ishandled beforehand.)
B. Cluster Labeling
After determining groups of users with similar music pref-
erences in each interval, we label each cluster with the most
often used genre tags to characterize the cluster. For this, we
combine the genre vectors of the user profiles in the cluster.
Definition 5: A cluster label cl of cluster i is defined as
�cli =∑n
u=1 �pun
, (3)
where n is the number of users in cluster i.Each detected cluster is thus labeled with a genre vector just
as a user profile.
IV. EXPERIMENTS AND RESULTS
In the following section, we briefly present the data set and
its characteristics. Furthermore, we discuss how we determined
the parameters ε and η for the clustering procedure anddescribe the results obtained by applying the incremental
DENGRAPH algorithm.
A. The Data Set
From the Last.fm Website we obtained the user listeningbehavior of 1,800 users over an interval of 130 weeks (from
March 2005 to May 2008). Last.fm provides for each userand interval (here one week) a list of the most listened to
artists and the number of times the artist was played. Based
on this information we determine user profiles for each interval
according to Definition 2.
B. Determining the Parameters ε and η
Before building the similarity graph according to Definition
4, the parameters ε and η have to be determined. For this,we conduct experiments using data from 16 intervals with
400 parameter combinations regarding the number of clusters.
Our aim is to determine a parameter combination that yields
an average number of clusters. The number of clusters varies
between six and nine and the average is obtained with ε = 0.12and η = 7.The similarity graph consists of edges (px, py) where (1−
sim(px, py)) ≤ ε. Thus, two nodes are only connected with anedge, if their distance is lower than ε. By this we achieve that
245245
only users with similar music preferences are in each other’s
ε-neighborhood. (Note, that the similarity would be one if twousers have equal profiles and zero if their profile vectors are
orthogonal. Thus, by subtracting the similarity from one, we
obtain the distance between two users.)
C. Data Set Characteristics
In the following, we present characteristics of the similarity
graphs over all weeks. In Figure 3, the number of nodes and
edges per interval are shown. The number of nodes and edges
increases significantly in the first 60 intervals and decreases
slowly after interval 18/07. The increase in the beginning is
due to the fact that many of the observed users were not yet
active on the platform at this time. The slow decrease could be
explained by a rising diversity in the genres that the users listen
to since this would result in lower similarity values between
users.
In Figure 4, the number of nodes per cluster and the fraction
of noise nodes per interval are shown. The clustering reveals
in each interval one giant component and two to eight smaller
clusters. The large component consists user profiles with a high
weight for the genres ’indie’, ’indie rock’ and ’alternative’.
The percentage of unclustered nodes (noise) is constantly
high and varies between 0.63 and 0.75. This observation
stresses the need for a clustering procedure that is able to
remove noise objects in linear time such as DENGRAPH.
Fig. 3. Number of nodes and edges
D. Results
We applied DENGRAPH on the data set to analyze the
evolution of clusters. As discussed in the previous section,
we obtain on average four clusters in each week. Due to the
labeling of the clusters we observe four main clusters in our
data set: ’indie’, ’hip-hop’, ’metal’ and ’rock’.
In the experiments, several cluster transitions can be ob-
served. To illustrate the observations, we chose a period of
thirty weeks and depict the clustering results of six weeks
as shown schematically in Figure 5. In each week, five to
Fig. 4. Number of clustered nodes (top); Fraction of unclustered nodes(bottom)
seven clusters were detected. The graph visualizations of the
clusterings for four weeks of our example period are shown
in Figures 6 and 7.
One cluster was stable over all periods (cf. Cluster 1 in
Figure 5). The label of this cluster revealed that it consists
of people who listen mainly to artists tagged with ’indie’ and
’alternative’. The number of people in this cluster varies over
the observation period only slightly (around 500) as shown in
Figure 4. Another cluster that can be observed over a long
period is labeled with ’hip-hop’ (cf. Cluster 4 in Figure 5).
This cluster can be observed in most weeks and does never
overlap with any other cluster. An explanation could be that
the genre ’hip-hop’ is very unique and has a low similarity to
other observed genres.
a) Week 29/06: In the first week of the period, fourclusters are detected. The labels of these cluster can be seen
in the legend of Figure 5. The four clusters represent mainly
the four genres that we observe in our data set: Cluster 1 is
the ’indie’ cluster, cluster 4 the ’hip-hop’ cluster and cluster 2
the ’metal’ cluster. At this time, cluster 3 is tagged with both
genres ’rock’ and ’metal’. We will see at a later point (week
48/06) that this cluster finally splits to a ’rock’ and ’metal’
cluster.
b) Week 33/06: The four clusters detected in week 29/06are still observable and two new clusters are created: Cluster
5 labeled with the genres ’metal core’ and ’hard rock’ and
cluster 6 is labeled ’power metal’.
c) Week 37/06: Five of the six clusters are still observ-able, only cluster 6 with the tags ’power metal’ and ’heavy
metal’ has been removed. However, the cluster will show up
again in week 45/06 (cluster 9).
d) Week 40/06: After one month, clusters 2 and 5 havemerged to cluster 7. This transition can also be seen in the
screenshot of the visualization in Figure 6. Both clusters
were labeled with tags related to the genre ’metal’ and the
new cluster combines these labels. For example cluster 2 had
246246
Week29/06
Week33/06
Week37/06
Week40/06
Week45/06
Week48/06
Cluster Label
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
5
6
7
8
2
3
4
2
3
4
9
10
11
indie rock, alternative
melodic death metal, metal, death metal
progressive rock, progressive metal, rock, classic rock
hip-hop
metal core, hard rock, rock, emo, screamo, punk, metal
power metal, heavy metal, metal, symphonic metal
metal core, hard core, death metal, metal, melodic metal
electronic, ambient, idm, electronica, indie, chill out
power metal, metal, symphonic metal
progressive metal, progressive rock, metal, progressive
progressive rock, progressive metal, rock, metal
1
2
3
4
Cluster Evolution
11 cluster unchanged
new cluster
cluster removal
cluster merge
cluster split
5
6 ./.
1 1 1 1
3
4
3
4
7 7
4 4
10
8 8 8
9 9
4
3
2
1
5
2
7
3 11
Fig. 5. Genre Evolution
among others the label ’death metal’ and cluster 5 the label
’metal core’. The new cluster 7 has both labels. Furthermore,
in this week a new cluster has been detected (cluster 8). This
cluster is stable over the rest of our observation period and is
labeled among others with the tags ’electronic’, ’ambient’ and
’chill out’. The subgenres ’ambient’ and ’chill out’ indicate
that this cluster is closer to the genre ’indie’ than for example
to other subgenres of the electronic music genre such as
’house’ or ’techno’. This explains, why the cluster has in week
45/06 an overlap with cluster 1, the ’indie’ cluster (see Figure
7(a)).
e) Week 45/06: In week 45/06, a new (old) cluster hasbeen (re-)established (cluster 9). This is the ’power metal’
cluster that we already observed in week 33/06. Besides this
creation, the clustering remains unchanged.
f) Week 48/06: In the next month, the previously men-tioned split occurs. Cluster 3 which was labeled with both
’rock’ and ’metal’ tags is split into the clusters 10 and 11.
Cluster 10 is labeled with ’metal’-related tags and cluster 11
with ’rock’-related tags. Both clusters are still overlapping
since they have members in common. Interestingly, we can
also see that the ’rock’-genre has a higher similarity to the
’indie’-genre than the genre ’metal’: Members of cluster 11
(’rock’) are connected to members that again have neighbors
in the ’indie’-cluster, while there is no connection between
members of cluster 10 (’metal’) to members in cluster 1
(’indie’).
2
4
3
5
1
(a) Clustering in Week 37/06.
3
1
4
8
7
(b) Clustering in Week 40/06.
Fig. 6. Cluster 2 and 5 in (a) merge to cluster 7 in (b) and a new cluster iscreated (cluster 8 in (b)).
g) : Our analysis showed that DENGRAPH detects
groups of users with similar music preferences by clustering
their profile. We obtain groups of users with a similar music
listening behavior. Some of these groups overlap with each
other and others are separated over the entire period of
observation. Clusters that overlap are more similar, because
they have tags in common. On the other hand, we detect groups
that are separated, because they have no similarity with others,
such as the ’hip-hop’ cluster.
Furthermore, we could show that the incremental procedure
allows to track and observe temporal dynamics of user clusters.
We observe the growth and decline of clusters as well as the
creation of new clusters and the removal of formerly existing
clusters.
Using DENGRAPH, we can also observe the evolution
of genre clusters: Transitions such as merges and splits are
247247
8
4
93
1
7
(a) Clustering in Week 45/06.
9
8
4
1
7
11
10
(b) Clustering in Week 48/06.
Fig. 7. Cluster 3 in (a) is split into cluster 10 and 11 in (b).
detected and due to the cluster labeling we are able to verify
that these transitions are semantically reasonable. For example,
we observe that a cluster consisting of members listening to
music tagged with ’rock’ and ’metal’ splits at some point into
two separate clusters, while one has the label ’rock’ and the
other ’metal’. On the other hand, we observe the merge of
clusters with similar tags.
V. CONCLUSION
In this paper we presented an application of the DEN-
GRAPH clustering algorithm on a data set obtained from the
social networking site Last.fm. Over a period of 130 weeks, wedetermined for each interval user profiles that represent music
preferences. By building and incrementally clustering a graph
of similar users for each interval, we obtain groups of users
with similar music preferences. We label all clusters according
to the user profiles of the members of the respective cluster.
Due to the incremental nature of DENGRAPH we could show,
how clusters evolve over time. We observe how clusters grow
and shrink, how new clusters are born and old clusters die.
In all these cases we notice that the listening behavior of
people has changed and that these changes result in a joining
or leaving of communities. Furthermore, we showed the merge
and split of clusters. By using the labels of clusters, we could
verify that the observed merges and splits are semantically
explainable since in both cases clusters with similar labels
were involved.
The experiments indicate that DENGRAPH is a clustering
algorithm capable to efficiently detect and track communities
in large, noisy graph structures. This is an important advantage
over, for example, widely used hierarchical clustering tech-
niques (see, e.g., [15]) as they do not scale to large social
networks. Due to its incremental update function, clusters
are observable over time and the temporal dynamics in these
networks can be revealed.
REFERENCES
[1] Eurostat, “Cultural statistics, http://epp.eurostat.ec.europa.eu/,(retrieved 23/08/2008),” 2007. [Online]. Available:http://epp.eurostat.ec.europa.eu/
[2] G. Kreutz, U. Ott, D. Teichmann, P. Osawa, and D. Vaitl, “Using musicto induce emotions: Influences of musical preference and absorption,”Psychology of Music, vol. 36, no. 1, pp. 101–126, 2008. [Online].Available: http://pom.sagepub.com/cgi/content/abstract/36/1/101
[3] P. J. Rentfrow and S. D. Gosling, “The content and validityof music-genre stereotypes among college students,” Psychology ofMusic, vol. 35, no. 2, pp. 306–326, 2007. [Online]. Available:http://pom.sagepub.com/cgi/content/abstract/35/2/306
[4] T. Hays and V. Minichiello, “The meaning of music in thelives of older people: a qualitative study,” Psychology ofMusic, vol. 33, no. 4, pp. 437–451, 2005. [Online]. Available:http://pom.sagepub.com/cgi/content/abstract/33/4/437
[5] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-basedcollaborative filtering recommendation algorithms,” in WWW ’01:Proceedings of the 10th international conference on World Wide Web.New York, NY, USA: ACM, 2001, pp. 285–295. [Online]. Available:http://portal.acm.org/citation.cfm?id=372071
[6] T. Falkowski, A. Barth, and M. Spiliopoulou, “Studying communitydynamics with an incremental graph mining algorithm,” in Proc. ofthe 14 th Americas Conference on Information Systems (AMCIS 2008),2008.
[7] J.-J. Aucouturier and F. Pachet, “Representing musical genre: A state ofthe art,” Journal of New Music Research, vol. 32, pp. 83–93, 2003.
[8] M. Schedl, T. Pohle, P. Knees, and G. Widmer, “Assigning and visualiz-ing music genres by web-based co-occurrence analysis,” in Proceedingsof the Seventh International Conference on Music Information Retrieval(ISMIR’06), Victoria, Canada, October 2006.
[9] R. Basili, A. Serafini, and A. Stellato, “Classification of musical genre:a machine learning approach.” in ISMIR, 2004.
[10] M. Levy and M. Sandler, “A semantic space for music derived fromsocial tags,” in 8th International Conference on Music InformationRetrieval (ISMIR 2007), 2007.
[11] C. McKay and I. Fujinaga, “Musical genre classification: Is it worthpursuing and how can it be improved?” in ISMIR, 2006, pp. 101–106.
[12] G. Geleijnse, M. Schedl, and P. Knees, “The quest for ground truth inmusical artist tagging in the social web era,” in Proceedings of the EighthInternational Conference on Music Information Retrieval (ISMIR’07),S. Dixon, D. Bainbridge, and R. Typke, Eds., Vienna, Austria, September2007, pp. 525 – 530.
[13] M. Ester, H.-P. Kriegel, J. Sander, M. Wimmer, and X. Xu, “Incrementalclustering for mining in a data warehouse environment,” in Proc. of 24thVLDB Conference, 1998.
[14] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithmfor discovering clusters in large spatial databases with noise,” in Proc. of2nd International Conference on Knowledge Discovery and Data Mining(KDD-96), 1996, pp. 226–231.
[15] M. E. J. Newman and M. Girvan, “Finding and evaluating communitystructure in networks,” Physical Review, vol. E 69, no. 026113, 2004.
248248