See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/229017694

Using virtual worlds for behaviour clusteringbased analysis

Conference Paper October 2010
DOI: 10.1145/1878083.1878088

CITATIONS

READS

39

4 authors:
Antonio Gonzalez-Pardo

Francisco de Borja Rodrguez Ortiz

Universidad Autnoma de Madrid

Universidad Autnoma de Madrid

31 PUBLICATIONS 113 CITATIONS

111 PUBLICATIONS 673 CITATIONS

SEE PROFILE

SEE PROFILE

Estrella Pulido

David Camacho

Universidad Autnoma de Madrid

Universidad Autnoma de Madrid

52 PUBLICATIONS 472 CITATIONS

171 PUBLICATIONS 693 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,

letting you access and read them immediately.

SEE PROFILE

Available from: Antonio Gonzalez-Pardo

Retrieved on: 20 June 2016

Using Virtual Worlds for Behaviour Clustering-based

Analysis
Antonio Gonzlez-Pardo, Francisco B. Rodrguez, Estrella Pulido and David Camacho
Departamento de Ingeniera Informtica. Escuela Politcnica Superior. Universidad Autnoma
C/Francisco Toms y Valiente 11, 28049
Madrid, Spain

{antonio.gonzalez, f.rodriguez, estrella.pulido, david.camacho}@uam.es

ABSTRACT

The attractive 3D graphical environments provided by

these VW can be used to improve interaction and a sense
of realism, for the users who access them. In these environments it is not only possible to see, hear and touch virtual
objects, but also to create, edit and manipulate them as if
they were physical objects.

Virtual Worlds have become a very popular domain and its

high inmersive characteristics can be used to extract information about the avatars behaviour. In this kind of environment it is possible to obtain interesting data about avatars,
such as their exact position in the world, what they are looking at (eye-gazing) or what they are talking about. This
paper studies how this information, obtained from avatars
interactions, can be integrated in order to apply clustering
techniques. Monitoring avatars in a virtual world is a useful
task that allows the identication of behavioral groups. The
meaning of these groups depends on the application domain,
for example in educational virtual worlds, they can represent
whether students are paying attention to the teachers explanation or not.

Although VWs have been used in dierent domains such

as Economy [12] or E-Commerce [13], the most popular
are related to massively-multiplayer on-line games, such as
World of Warcraft or Doom [1]. VWs can be used as a new
powerful instrument for instruction and education, where
characteristics like persistence allow continuous and growing
social interactions, which can be the basis for collaborative
education [9, 11]. However, there are some characteristics
that have not been fully explored in this kind of environments. Due to the immersive characteristics of VWs and the
possibility to acquire automatically complex data from users
(avatars) such as their current position, eye-gazing data, and
other kind of interactions such as conversations, or operations performed in the virtual world (building objects, results of script compiling, etc...), it is possible to design real
experiments where the behaviour of a set of avatars can be
analyzed. Some of the previously described features could
be extracted and analyzed in a real environment but with
a very high complexity (i.e. to obtain the eye-gazing information from a set of human users very complex vision techniques would need to be used with its consequent high cost).

Categories and Subject Descriptors

I.5.3 [Pattern Recognition]: ClusteringAlgorithms, Similarity measures

General Terms
Algorithms, Experimentation

Keywords
Virtual Worlds, Hierarchical Clustering, Normalized Compression Distance

1.

This paper uses a new VW platform, named V-LeaF, that

allows the design and implementation of virtual spaces where
a set of avatars can be intensively monitorized by using a set
of tools that can be managed by the platform administrator
or by a particular avatar (i.e. the teacher in an educational
domain). The paper also shows how clustering techniques
and information theory-based techniques can be used to analyze the dierent actions, and behaviours, shown by avatars
during their stay in the virtual world. The Normalized Compression Distance (NCD) metric is used to calculate the similarity between dierent items (i.e. text les). This metric
is combined with a hierarchical clustering to analyse the
avatars behaviour.

INTRODUCTION

Virtual Worlds (VW) have become a very popular kind of

software application that has been used in dierent elds,
from games to simulation or education. They allow individuals to interact with others through their avatars and
with objects in the environment. This interaction and its
3D appearance makes Virtual Worlds a good environment
to develop educational and collaborative tasks.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SMVC10, October 29, 2010, Firenze, Italy.
Copyright 2010 ACM 978-1-4503-0175-6/10/10 ...$10.00.

(a) First Stage of a lecture on Ancient Greece

(b) Second Stage of a lecture on Ancient Greece

Figure 1: Screenshots of the dierent stages of a lecture on Greek history. At a rst stage (a), the teacher
explains dierent concepts about Ancient Greece. At a second stage (b) students are collaboratively working
to build a classical monument related to the concepts explained at stage (a).

2.

V-LEAF: A VIRTUAL WORLD

PLATFORM

3. BASICS ON NORMALIZED COMPRESSION DISTANCE

V-LeaF (Virtual LEArning platForm)1 [9, 10] is a VW

platform built on a computer grid over the OpenSim [8] simulator. This platform provides some facilities and software
tools that allow to automatically acquire data from avatars
actions in the VW. Although this platform was initially designed with an educational perspective, it has been extended
to be used in other application domains such us medicine or
bioinformatics. The main software characteristics of V-LeaF
can be briey summarized as follows:

In this paper, the Normalized Compression Distance (NCD)

metric [3, 6] is used to compute the similarity distance between the students chat conversations. The use of NCD
with text les to perform clustering tasks has been widely
used and applied to dierent problems and domains as shown
in [3, 4, 6].
The NCD [3, 6] provides a measure of the similarity between two objects, x and y, by using compressors. The
formal denition of NCD is shown in equation 1.

A VW grid that provides virtual spaces where lectures

and laboratory activities can take place and that allow
the storage of objects created by dierent avatars.

N CD(x, y) =

(1)

where C is a compression algorithm, C(x) is the size of

the C-compressed version of x, and C(xy) is the compressed
size of the concatenation of x and y. NCD generates a nonnegative number 0 N CD(x, y) 1. Distances near 0
indicate similarity between objects, while distances near 1
reveal dissimilarity.

A back-end service built on the WampServer web development environment and a SQLite database engine
designed to store technical and user guides, educational documentation such as courses guides, multimedia, group information (avatar proles, passwords,. . . ),
and a set of tools that allow the storage of dierent
logs containing VW data (position, eye-gaze, chat,. . . )
extracted from avatars interactions in the VW. This
back-end service provides an interface to automatically
map data logs to an adequate format for data analysis
and statistical software tools (i.e. it is possible to select dierent features extracted from the logs and map
them into .ar format for the Weka toolkit, or to be
processed by using Matlab).

The NCD is just one of the many similarity distances that

use compression algorithms. Others [2, 5, 15] are small variations and can be easily reduced to it, since it is possible
to prove that this distance minorises (it is as good as) any
other that can be computed by a universal Turing machine
[14].

4. A LECTURE ON HISTORY

A Web portal that provides access to public documentation (courses, technical and user guides,. . . ) and a
user/group management service that allows for three
dierent roles in the platform (administrator, group
responsible (i.e. educators) and regular users (i.e. students). This service allows administrators to manage
and control the regular users enrolled in a experiment
(i.e. course).
1

max{C(xy) C(x), C(yx) C(y)}

,
max{C(x), C(y)}

As it is described in Section 2, by using the OpenSim Simulator, the V-LeaF platform oers a new environment where
dierent kind of islands (or simply sims) can be built for
dierent goals. Currently V-LeaF has deployed several islands related to educational topics and a medicine island. VLeaF has been widely used in cooperation with High School
institutions to provide a controlled access to educational
spaces. Each high school has its own island where teachers can control any educational activity that they design

http://this.ii.uam.es/vleaf/en/index en.jsp

10

(a)

(b)

Figure 2: Representation of the eye-gazing data. Figure (a) shows how the avatar view is not the same as
the user view. Figure (b) shows the same scene but from a dierent perspective. The angle formed between
the avatar and user view, called eye-gazing is represented as a red arrow.
public channel to give a lecture and communicate with all
the students at the same time.

(such as lectures, demonstrations, collaborative experiences,

etc. . . ) [10]. One of those islands is dedicated to history education. The data used in this paper was extracted from a
lecture on history.

The private channel is the other way of communication

within the VW. By using this channel a bidirectional communication is created between two parts. This communication channel is a point to point communication, which means
that nobody else will receive the message except the receiver.

In this study, a lecture was given in a Greek amphitheatre where the teacher explained some basic concepts about
ancient history, ancient Greek art, monuments and temples.
During the rst stage of the lecture, all the students should
be paying attention to the teachers explanation. As Figure
1 shows, avatars are in the amphitheatre and some of them
are looking at the teacher. This stage is very important because in the second part of the class, students will work in
groups to build some historical buildings proposed by the
teacher.

From the clustering point of view, messages sent through

both channels are relevant, because all the conversations are
useful to determine if two or more avatars are talking about
the same topics and whether this topic is related to the lecture. For that reason, all messages sent through the public
or private channels are stored in order to be analyzed.
An avatar position is represented by the Cartesian Coordinates System, which dene the location in the VW. As
a rst approach, position can be used to determine what
avatars are paying attention to simply by dening a scope.
For example, a teacher could consider that if the distance
between the avatar and the teacher is greater than X meters,
the student is not paying attention. This approach matches
perfectly with the three types of messages that any avatar
can send through the public chat.

Avatars will be grouped depending on its preferences and

at the end of the lecture, the teacher will evaluate their
buildings. Figure 1 shows how avatars are working in their
dierent tasks in this second stage. The students must use
basic geometric primitives such as pyramids, cylinders or
cubes, in order to build a historical monument.

5.

THE DATA ACQUISITION PROCESS

Finally, the eye-gaze is a calculated parameter that depends on two factors. On the one hand, it depends on the
distance from the avatar to the teacher. On the other hand,
it will depend on the user viewpoint or the user camera orientation. It is important to take into account that in virtual
worlds, there are two types of views. The rst one is the
avatar view which is what the avatar should see. And the
other view is the user view, or camera view, which is what
the real user is seeing on the screen. This last view is the
one used to compute the eye-gazing, because the purpose is
to determine when the real user is paying attention to the

In V-Leaf, three kinds of dierent data are recorded from

each avatar which are chat, position and gaze. The chat
stores what the avatar says in the virtual world. There are
two channels through which an avatar could send a message:
the public and the private channel. By using the rst one
any avatar near the sender will receive the message. The
meaning of near will depend on the type of message sent
because an avatar could whisper a message (in this case any
avatar within a radius of 10 meters will receive the message),
say a message within 20 meters, or shout a message until 100
meters. In the educational domain, the lecturer will use the

11

variable used to perform the clustering is eye-gaze which

means that the task of determining whether an avatar is paying attention to the teacher depends exclusively on the gaze
of the user and not on the distance from the avatar to the
teacher. By analyzing all data collected at the rst phase,
results show that the angle that describes the behaviour of
the user is 36.45 degrees. This means, that if the gaze of a
user is greater than this value, the user is not paying attention to the explanation given by the teacher. Nevertheless,
this result is extremely dependent on the dataset used in the
experiments. In order to determine the information that denes whether a student is paying attention or not, a more
complete study must be performed.

class and not whether the avatar is paying attention or not.

With these concepts the eye-gaze is compute as follows:

camera
separation
(2)
EyeGaze = arccos(
)
|camera| |separation|
is the vector dened by the user view and
Where
camera

separation is the imaginary vector that connects the avatar

with the teacher. The eye-gaze is represented in Figure 4

and Figure 4. The avatar view (separation) is represented

in Figure 4
by a blue arrow, and the user view (
camera)
is represented by a black line. The eye-gaze represents the

and
angle formed by
camera
distance, and it is represented
by a red arrow.

150

Position and gaze are recorded in three second intervals

and this information is saved with a timestamp to facilitate
its analysis. The chat is recorded when the user talks with
other avatars. The timestamp for chat conversations corresponds to when the conversation was initiated.
EyeGaze

6.

100

EXPERIMENTAL RESULTS

The experiment carried out in this work involves nine

avatars (eight students and a teacher) and was divided into
two phases. The rst one took an hour and the other one
took half an hour.

50

During the rst phase of the experiment, the teacher was

explaining some concepts about Ancient Greece. Among the
data stored in this phase, the chat information is the less relevant. This is an expected situation because if students are
paying attention to the lecturer, they should not use intensively the chat. For this reason, to determine whether a
particular avatar is paying attention or not to the lesson,
only its position and gaze data are used and the K-means
algorithm [7] is applied to them. This algorithm will divide
the set of recorded data into two classes corresponding to
the students that are paying attention and those who are
not.

0
0

Distance

10

15

Figure 3: Representation of the avatars angle of vision (y-axe) and their distance to the teacher (xaxe), before the lecturer requires their attention.

60
50

Figures 3 and 4 represent information about avatars. Each

asterisk represents an avatar in a specic time. The information specied for each avatar includes the distance
from the avatar to the lecturer and the angle formed by the
user-camera view and the distance just explained (eye-gaze).
For example, in Figure 3, the circled asterisk represents an
avatar that has a distance of 9 meters to the lecturer and an
eye-gaze angle of around 100 degrees.

EyeGaze

40
30
20

Figure 3 shows distance from the avatars to the lecturer

and the avatars eye-gaze before the lecturer requires their
attention and Figure 4 represents the same data but when
the lecture has already started. As it can be seen, values for
distance and eye-gaze when the lecture has started are more
compacted than in Figure 3. This is an expected result because when the teacher is giving explanations, students are,
more or less, within the same distance to the teacher.

10
0
0

6
8
Distance

10

12

14

Figure 4: Result of the K-means algorithm applied

to avatars angle of vision (y-axe) and their distance
to the teacher (x-axe) during the rst stage of the
lecture.

Figure 4 shows an example of the result of K-means over

the data taken during the lecture. It can be seen that the

12

 Group 3. Student 7 and Student 8. Teacher requires

both students to build the Temple of Delians in Delos.

Taking the previous limit for the eye-gazing, the task of

paying attention can be dened as whether the eye-gaze angle is less than 36.45 degrees. With this value, the eye-gazing
of each avatar, for example student X, has been divided into
two les depending on whether the student is paying attention (YstuX ) or not (NstuX ). Then, a hierarchical clustering
using NCD has been performed over both set of les in order to observe the behaviour of NCD-based clustering over
those les. Figure 5 shows the results of the clustering process. This new test shows how most of these les has been
correctly classied according to the avatars behaviour, and
how both clusters are clearly dened. Therefore, the NCD
can be directly used over eye-gaze data only extracting the
numerical values from avatars interactions and using this
metric to measure the attention level.

Figure 6 shows the NCD-based clustering results over the

chat conversation. It can be seen that the three groups are
perfectly identied in the dendrogram. Group 3 is the closest
group to the teacher, because its components talked many
times about the Doric, Ionic and Corinthian styles. Moreover, Student 3 is near Group 3 because he explained how
to build basic objects to that group.
Members belonging to Group 2 are grouped together and
they are further away than Group 3 because Group 2 did
not talk about the concepts the explained by the teacher.

0.908

0.696
0.949

Yst2

0.947

student1

Yst8

Yst5

0.852

teacher

student2

student8
0.819

0.950

0.941

0.925

0.801

0.913
Yst6

Yst7

Nst4

student4
Nst1

0.944

student7
0.919

0.737

Yst1

student3

Nst5

0.943

0.895

Yst4

0.797

student5
student6

Nst7

0.846

0.719

0.929
Nst6

S(T)=0.997067

Nst3
0.890
0.921

0.945

Figure 6: Dendrogram resulting from the clustering

of the messages exchanged between students using
NCD which shows that groups belong to the dierent user proles.

Yst3
Nst8

Nst2

0.906

0.890
S(T)=0.994739

7. CONCLUSIONS

Figure 5: Results of the NCD-based clustering over

the les containing the eye-gaze of each avatar. The
rst letter of the name indicates whether the student
is paying attention (Y) or not (N).

Virtual Worlds have been used from games to business or

educational purposes. Some of the main work carried out in
this eld is directly related to the utilization of this highly
immersive domain to study social behaviour, and cooperative and collaborative skills from individuals. However, no
much eort has been carried out to integrate the available
information such as the avatar position, eye-gazing data or
textual information exchanged in chat conversations to use
data mining techniques, such as clustering, to automatically
analyze the avatar behaviour.

In the second phase, the representation of the distance

and the gaze of the avatar are not relevant because avatars
work in groups to build the objects explained in the rst
phase. Nevertheless, in this phase there was an important
increase in the usage of the chat because avatars talk to their
classmates in order to build the required objects correctly
or to distribute the dierent tasks among team members. In
this phase students are grouped as follows:

This paper shows an initial result in this direction, where

data extracted from an educational experiment has been
processed and integrated to detect the behaviour of a set
of students involved in some educational tasks. The experimental results show that integration of dierent data is
useful to classify students into several groups such as students paying attention and distracted students. The use
of a K-means algorithm over the set of gazing and position
data reveals that eye-gazing is the most important feature
to determine whether the student is paying attention. This

Group 1. Student 1 and Student 2. This group does

not build any object because both students spend their
time exploring and trying some features (chat, teleport, ying,...) in the virtual world.
Group 2. Student 3, Student 4, Student 5 and Student 6. This group is made up of students with a high
knowledge on 3D design. Their task was to build the
Parthenon.

13

[5] A. Kraskov, H. Stoegbauer, R. Andrzejak, and

P. Grassberger. Hierarchical clustering using mutual
information. Europhysics Letters, 70(2):278284, 2005.
[6] M. Li, X. Chen, X. Li, B. Ma, and P. Vitanyi. The
similarity metric. IEEE Transactions on Information
Theory, 50(12):32503264, 2004.
[7] J. B. MacQueen. Some methods for classication and
analysis of multivariate observations. In L. M. L. Cam
and J. Neyman, editors, Proc. of the fth Berkeley
Symposium on Mathematical Statistics and
Probability, volume 1, pages 281297. University of
California Press, 1967.
[8] [Online] Available:
http://opensimulator.org/wiki/Main Page.
OpenSimulator.
[9] M. Rico, D. Camacho, and E. Pulido. Automatic
guidance tools for enhancing the educative experience
in non-immersive virtual worlds. In Proc. of the IEEE
Engineering Education Conference-EDUCON 2010,
volume 1. IEEE Press, 2010.
[10] M. Rico, G. Martnez-Mu
noz, X. Alaman,
D. Camacho, and E. Pulido. A programming
experience of high school students in a virtual world
platform. International Journal of Engineering
Education, page In press, 2010.
[11] T. Ritzema and B. Harris. The use of second life for
distance education. J. Comput. Small Coll.,
23(6):110116, 2008.
[12] E. Sinrod. Virtual world litigation for real, 2007.
[13] D. Talbot. The eecing of the avatars. Technology
Review, 111(1):5862, 2008.
[14] A. Turing. On computable numbers, with an
application to the entscheidungsproblem. Proceedings
of the London Mathematical Society, 2(42):230265,
1936.
[15] X. Zhang, Y. Hao, X. Zhu, and M. Li. Information
distance from a question to an answer. In KDD 07:
Proceedings of the 13th ACM SIGKDD international
conference on Knowledge discovery and data mining,
pages 874883, New York, NY, USA, 2007. ACM.

means that the distance from the teacher to the avatar is not
relevant when the student is looking at the teacher. Text exchanged through the chat can be used to classify students.
In this case, the lecture needs to be analyzed in order to
determine which student is talking about the concepts explained by the teacher. A hierarchical clustering over the
similarity metric of the chat record, shows the groups of
students depending on the contents of their conversations.
Nevertheless, a deeper study on the integration of dierent types of data needs to be performed to determine the
basic rules that dene whether a student is paying attention
or not. These rules could be used to generate recommendations to the teacher about the students that are not paying
attention to his/her explanations and to help him/her taking the best decision about the next action.

8.

ACKNOWLEDGMENTS

This work has been funded by the Spanish Ministry of

Science and Technology under the projects TIN2007-65989,
TIN2007-64718 and TIN2008-02729-E/TIN. We would like
to acknowledge Victor Cabezas, Daniel Bravo and Daniel
Gomez for their work in the implementation of the V-LeaF
platform.

9.

REFERENCES

[1] F. Bellotti, R. Berta, A. D. Gloria, and V. Zappi.

Exploring gaming mechanisms to enhance knowledge
acquisition in virtual worlds. In DIMEA 08:
Proceedings of the 3rd international conference on
Digital Interactive Media in Entertainment and Arts,
pages 7784. ACM, 2008.
[2] D. Benedetto, E. Caglioti, and V. Loreto. Language
Trees and Zipping. Physical Review Letters,
88(4):48702, 2002.
[3] R. Cilibrasi and P. Vitanyi. Clustering by
compression. IEEE Transactions on Information
Theory, 51(4):15231545, 2005.
[4] A. Granados, M. Cebri
an, D. Camacho, and F. B.
Rodrguez. Reducing the lost of information through
annealing textual distortion. IEEE Transactions on
Knowledge and Data Engineering, page To appear,
2010.

14