The Design of Gugubarra 2.0:
A Tool for Building and Managing Profiles of Web Users
Natascha Hoebel, Sascha Kaufmann, Karsten Tolle, Roberto V. Zicari
Databases and Information Systems (DBIS)
Johann Wolfgang Goethe-University of Frankfurt
Robert-Mayer-Strasse 11-15, D-60325 Frankfurt, Germany
{hoebel, kaufmann, tolle}@dbis.cs.uni-frankfurt.de
zicari@informatik.uni-frankfurt.de

Abstract
In [6] we have introduced the concept of non-obvious
user profiles (NOPs) to capture the hypothetical interest
of web users. In this paper we present the design
principles and rules of our Gugubarra engine, which is a
tool to calculate and visualize these non-obvious user
profiles.

1. Introduction
With our Gugubarra project (Gugubarra is the
Aboriginal name for the Kookaburra bird) we try to build
user profiles representing the users’ interests. User
profiles can be used for recommender systems, which is
not our main objective. Goals we have in mind are as
follows:
• clustering of user groups by interests,
• detect shifts in interests of users or user groups,
• detect upcoming trends of the community.
Reaching these goals would enable the owner of a
web site to understand the community that is visiting his
site. He therefore could concentrate for example his
efforts on those areas that are potentially of interest to the
community. He could also get a feedback if his work is
adopted as he anticipated.
The Gugubarra project began in 2004 within the
database group (DBIS) at the Computer Science Institute
of the Johann Wolfgang Goethe University, with the aim
to build tools for better management of communities of
registered Web visitors.
A first research prototype system, called Gugubarra
1.0 has been implemented in 2004, which allows to build
and manipulate non-obvious user profiles [6]. It was
showcased at the CeBIT Trade Fairs in 2005 and 2006
[1]. Gugubarra 1.0 works as a test application on real data
provided by the Web community viewzone.org.
We are currently designing the second prototype
called Gugubarra 2.0, where we offer various settings that

can be used to influence the calculation of the profiles. By
using these settings the Web site owner can focus on
aspects he wants to analyze and adjust the tool according
to his community or target group.
The design principle of the Gugubarra engine is as
follows: NOPs are calculated by using several different
parameters that can be chosen by the owner of a Web
portal. The combination of the chosen parameters
implements a specific strategy to deploy and manage
NOPs. The common element for all strategies that can be
chosen is as follows: The Creation of NOPs is done by
looking at the behavior of the web user and by taking into
account a feedback mechanism. More precise the
approach we have in Gugubarra is as follows:
i.

ii.

iii.

For each registered Web visitor we create a
profile. These user profiles reflect the “inferred”
interests of the users related to a set of predefined topics defined by the owner of the Web
site. The profiles go beyond collecting the
obvious information the user is willing to give at
the time of registration. In Gugubarra, a user
profile contains two parts: the obvious profile,
given directly by the user and a non obvious
profile (NOP), inferred by the user’s behavior
during his visits on the site.
The user profile is (re)-calculated dynamically
every time an explicit feedback is given by the
user and/or a set of events occurred which are
related to the user’s behavior and to certain
“locations” of the Web site.
We cluster Web visitors by clustering similar
profiles of interest [4]. Cluster of Web visitors
can then be used to analyze patterns of interests
in the Web community and to forecast further
behavior. Clusters might also provide useful
information to support the decision what kind of
new E-services to introduce for the Web
community and when to introduce them.

1

In this paper we’ll give an overview of the differences
between Gugubarra 1.0 and the new version. The effect
of the various parameters, in defining a strategy to create
the NOPs, will also be emphasized.

2. Gugubarra 1.0 Review
The approach to calculate the NOPs with Gugubarra
1.0 was only page and duration based. The strategy has
been defined detailed in [6]. Basically it consists of
associating each page Pj of the web site to a set of predefined topics Tp1,..,Tpn, which are weighted statically. A
weight vj(Tpi), (0 ≤ vj(Tpi) ≤ 1) is defined for each page,
representing the strength of the relation between the page
Pj and the topic Tpi.
Each time a user looks at a page Pj, the corresponding
set of topics is considered depending on the associated
weights. The initial NOP is then created by setting a value
xi, between 0 (no interest) to 1 (strong interest) to the
associated topic list by considering the pages the user has
looked at during a session. The values xi of the NOP are
calculated using the following formula, where
duration(Pj) takes into account the time spent by the user
on the specific page (see Formula 1).
m

xi =

∑ duration( P ) * v
j

j =1

j

m

(Tp i )

∑ duration( P )
k =1

(1)

k

We discovered some issues that limited the approach
of Gugubarra 1.0. First of all it is page based. One page
can contain various parts with different content. Normally
you do not know which part a user views inside his
browser. However, in case he presses a link or enters
some information, e.g. into a form, you could take these
action of the user into account.
This should be reflected within the calculation. In
addition many sites are generated by content management
systems that build up their pages out of different smaller
parts. As consequence we need to break down the pages
to smaller units, which we will call zones.
In Version 1.0 only the duration is used. In addition
the actions of the users should also be taken into account.
Note that the duration is somehow tricky since you will
never know what exactly the user is doing in front of the
screen, if he stays there at all. In the next section we
introduce zones, actions and other concepts we use with
Gugubarra 2.0.

3. Concepts for the NOP calculation in
Gugubarra 2.0
We consider a Web site (or a domain) as a collection
of web pages that are linked together within the site. Each
page has specific content.
In Gugubarra 2.0 the following main concepts are
introduced to create a user’s NOP: Zones with a state,
Topics, Weights, Actions and Feedback. There are
several different parameters in Gugubarra 2.0 that
combined determine a specific strategy to deploy and
manage NOPs and influence the usage of the concepts.
Some of these possible settings are described within the
next section. Here we give a brief description of the main
concepts.
A zone defines an area on the Web site. It can be a set
of pages, a set of parts within a page, a set of parts of
several pages, or any combination thereof.
A zone has one of three states. The state ON indicates
that this zone is being used to calculate the NOPs of the
visitors, state OFF indicates the zone is not used. In the
third state OFF-ACTION-SENSITIVE, only actions the
user does within the zone are taken into account (formula
(3)), not the duration (formula (4)), which will be
described in the next section.
Topics are related to the content and are defined global
by the owner of the Web site and then associated to the
zones. A weight indicates the relative importance of the
topic in respect to a scale from 0 (not relevant) to 1
(extremely relevant) in the zone.
Actions are also global defined by the owner of the
Web site, so they are applicable to any zone defined for
the Web site. Each action has an associated weight which
indicates the importance given to such action by the
owner of the Web site, ranging from 0 as minimum up to
n as max.
The feedback is an integral part of the NOP approach.
This mechanism is triggert due to certain criteria, where
upon the user is requested to give his feedback. Here the
user explicitly answers questions to determine his
interests. We use the given feedback to measure the
accuracy of the calculated NOP. However, we can not
assume that all users will give correct answers, e.g., on
purpose, due to lack of time or because of a
misunderstanding. The influence on the NOP of the user
is therefore limited by certain rules [3][4][6].
By the way, we believe that it is important that the user
is able to see and determine his profile. We know that
there are some difficulties relating to this approach, e.g.
the possibility of the user to influence his profile.
However, we think this is important for the user to feel
comfortable and to trust the system.

2

4. Calculating the NOP

j

In this section we describe the general way of
calculating the NOP.
In general the NOP is a set of values representing the
level of interest for a set of topics. The calculations of
these levels of interest (see formula (2) below) are
determined by two parts:
i.
Action Profile: the actions a user does in the
zones on the topic Tpi, which is computed in the
Action Profile ActP(i)
and
ii.
Duration Profile: the time, a user spends on
pages associated to the topic Tpi, which is
computed in the Duration Profile DurP(i).
These two parts can be parameterized by the Web site
owner in accordance to his needs. This means the owner
can increase or decrease the impact of the time a user
spends on a page compared to the actions he did.
It is worth mentioning that there are some problems
related to the time a user spends on the page, because you
will not know if the user is really reading this page or has
fallen asleep. Of course this can partially be absorbed by
using timeouts. However, by allowing the Web site owner
to adjust the parameters to the measured behavior of the
community, we expect better results and a higher level of
trust the Web site owner gives to the results.
In the overall formula (2), the parameters a and b are
used to customize the ratio between ActP(i) and DurP(i)
(implemented by a slider in Gugubarra 2.0).

xi = a * ActP(i ) + b * DurP(i ),
where (a + b) = 1

(2)

Let’s have a closer look at the two parts of the formula
(2), as defined by (3) and (4) below.
To compute ActP(i), we determine zone q, where the
action occurred and obtain the associated topic weight
v(Tpi, Zq). We then multiply this value by the sum of all
weights for all occurred actions in this zone. Finally we
calculate the sum of all zones, where an action occurred
and the associated topic lists contain topic i, divided by
the sum off all occurred action weights.
DurP(i) is computed in a similar way. We consider
each visited page Pj, that contains the topic Tpi and
multiply the time the visitor spent on this page
duration(Pj) by its topic weight v(Tpi, Pj). We finally sum
these values and divide it by the total time.

⎛

⎞
* v(Tp i , Z q ) ⎟
⎠
aw
∑ s

∑ ⎜⎝ ∑ aw

t

ActP(i ) =

q

t

s

∑ (duration( P ) * v(Tp , P ))

(3)

DurP(i ) =

j

i

∑ duration( P )

j

(4)

k

k

In fact, it isn’t easy to define the duration(Pj) used in
formula (4). What we want is the time the user viewed the
page. The problem is that from the log files we only know
when he requested the page and when his next request
was performed, in case there is a next request.
The Web site owner can define time limits that should be
taken into account for the duration profile part. We did
not include them into the formula (3) to keep it easier.
However, the basic idea is simple. The Web site owner
can define a minimum and maximum time for
duration(Pj). Everything out of this range will not be
considered.
In case a visitor visits a page less time than the
specified minimum, the system expects the requested
page did not contain the content the visitor expected.
Therefore the duration part will not be taken into account.
In case the visitor stays longer on the page than the
defined maximum time, the system limits the influence of
it, due to the fact that we do not know if the user is really
engrossed or if he has fallen asleep.
As an additional fact we take the design of the Web site
into account. The designer can decide whether the link
shall be opened inside the same window overwriting the
parent page or inside a new window. The later case is
normally done by the designer to allow the user to
perform additional actions on the more important parent
page. Therefore we keep on counting the duration for
parent page up to a defined value (by default the
maximum value). In case Gugubarra 2.0 realizes that
another action has been performed on the parent page the
duration is further increased. Note: We only take the
design view into account, of course the user is free to
open links in new windows.
For the minimum and maximum time we are currently
also working on an approach to take the amount of data
contained inside the page into account. Of course for a
page with very little content the minimum and maximum
time should be smaller than for a Web page displaying a
specification with more than 50 printed pages. Also the
average time a specific user needs to read a word on a
page could give an indication how to set the minimum
and maximum time (average reading speed, see [5]).
Note that in formula (4) we used the value v(Tpi, Pj) to
compute the duration part DurP(i). This should be the
weight of the topic Tpi within the page Pj.
In Gugubarra 2.0 however the owner associates topics
and their weights to zones and not to pages. To calculate
the NOP therefore we generate a page topic list composed
by the zone topic lists for each page. For this calculation
only zones with state=ON are taken into account. In
addition the Web site owner can decide individually for

3

every page how this topic list is generated by using one of
the following rules:
•

MAX-Rule: For each topic Tpi, scan all zones on a
page j and take the maximum topic weight.

v(Tpi , Pj ) = max(v(Tpi , Z q ))∀Z q ∈ Pj
•

MIN-Rule: For each topic Tpi, scan all zones on a
page j and take the minimum topic weight.

v(Tp i , Pj ) = min(v(Tp i , Z q ))∀Z q ∈ Pj
•

AVG-Rule: For each topic Tpi, scan all zones on a
page j and take the average topic weight.

v(Tp i , Pj ) = avg (v(Tp i , Z q ))∀Z q ∈ Pj
with Zq is a zone with state=ON.

5. Choosing the Parameters to define a
Strategy
Each community is different and behaves differently.
It is therefore important for the Web site owner to be able
to configure how the calculation of the NOPs for his
community will be performed and what will be taken into
account under which circumstances. For our next
prototype system Gugubarra 2.0 we therefore will give
users of the system the ability to define parameters that
will influence the NOP generation. Let’s revise the effects
of the parameter settings, done by the owner:
• Setting the topics and their weights for each zone.
This is fundamental and will affect both parts of
the formula (1).
• Define specific actions, the Web site visitors
could perform, together with their relevant
weights. This will influence the action profile part,
represented in formula (2).
• Turning the states of the zones ON or OFF. This
will influence the page topic list generation and
therefore influence the duration profile part
represented in formula (3).
• Using the different rules to calculate the page
topic list (MAX-Rule, MIN-Rule or AVG-Rule)
will influence the duration profile part.
• By specifying the factors a and b in formula (1),
the action and duration profile parts can be
balanced by the Web site owner. With the extreme
case of a=0 and b=1 actions are turned off in the
sense that formula (2) will not be considered for
the NOP generation. However, actions still might
have an influence by turning ON or OFF the state
of zones and therefore can influence the page
topic list generation.
• Defining the maximum and minimum duration
that will be considered for the duration part.

A peculiarity of Gugubarra is the combination of
NOPs, user feedback and custom rules for the generation
of profiles. This is missing in systems such as [10][11].
Compared to systems such as Web Usage Mining [7] we
included more granular information by the introduction of
zones and actions, and we do not only rely on the
recorded click stream like in [2], [8], [10]. A detailed
example of the use of Gugubarra 2.0 can be found in [3].

6. References
[1] DBIS J.W. Goethe University, CeBIT Presentations for
Gugubarra 1.0., http://www.dbis.informatik.unifrankfurt.de/news/?mode=art&l=e&aid=2&tmid=3&smid=7,
2006.
[2] Y. Fu, K. Sandhu, and M. Shih, Fast Clustering of Web
Users Based on Navigation Patterns, in World Multiconference
on Systemics, Cybernetics and Informatics (SCI/ISAS'99),
pages 560--567, Orlando, FL, 1999.
[3] Natascha Hoebel, Sascha Kaufmann, Karsten Tolle and
Roberto Zicari, The Gugubarra Project: Building and
Evaluating User Profiles for Vistors of Web Sites, First IEEE
Workshop on Hot Topics in Web Systems and Technologies,
Boston, Massachusetts, USA, November 2006.
[4] Natascha Hoebel and Roberto Zicari, On Clustering
Visitors of a Web Site by Behavior and Interests, DBIS report,
May 2006.
[5] T.-P. Liang and H.-J. Lai, Discovering User Interests from
Web Browsing Behavior: An Application to Internet News
Services, IEEE Computer Society, Los Alamitos, CA, USA,
2002.
[6] Naveed Mushtaq, Karsten Tolle, Peter Werner and
Roberto Zicari, Building and Evaluating Non-Obvious User
Profiles for Visitors of Web Sites, IEEE Conference on ECommerce Technology (CEC 04), San Diego, California, USA,
2004.
[7] J. Srivastava, R. Cooley, M. Deshpande, Pang-Ning Tan,
Web usage mining: discovery and applications of usage patterns
from Web data, ACM Press, New York, USA, 2000.
[8] Qing Wang, Dwight J. Makaroff, H. Keith Edwards,
Characterizing Customer Groups for an E-Commerce Website,
Proceedings of the 5th ACM Conference on Electronic
Commerce (EC '04). ACM Press, New York, 2004.
[9] Weinan Wang, Osmar R. Zaïane, Clustering Web Sessions
by Sequence Alignment, Proceedings of the 13th International
Workshop on Database and Expert Systems Applications, 2002.
[10] S. Acharyya, J. Ghosh, Context-Sensitive Modeling of
Web-Surfing Behaviour Using Concept Trees, WEBKDD, 2003.
[11] B. D’Ambrosio, E. Altendorf, J. Jorgensen, Probabilistic
Relational Models of On-line User Behavior, WEBKDD 2003,
2003.

4