ENHANCING RECOMMENDER SYSTEMS DEVELOPMENT
WITH HUMAN, SOCIAL, CULTURAL AND ORGANIZATIONAL
FACTORS
Andreas S. Andreou
Department of Electrical Engineering and Information Technology, Cyprus University of Technology
P.O.Box 50329, 3603 Limassol, Cyprus
Stephanos Mavromoustakos
Department of Computer Science and Engineering, European University Cyprus, P.O.Box 22006, 1516 Nicosia, Cyprus
Keywords: Web Recommender Systems, Human, Cultural, Social and organizational factors.
Abstract: Recommender Systems (RS) aim at suggesting filtered Web information adapted to the needs or interests of
users by predicting their access behavior using a certain strategy or algorithm. The creation of RS is usually
approached focusing mostly on user behavior modeling, while the recommendation engine often neglects
critical, non-technical aspects of software systems development. Conceiving a RS primarily as a self-
contained or part of a Web-application, the present paper utilizes the SpiderWeb methodology and takes
into account important requirements that result from human, cultural, social and organizational factors
(HSCO) so as to drive the RS development activities.
1 INTRODUCTION
Recommender Systems (RS) are a special class of
software applications usually running on the Web to
support users in making decisions while interacting
with large spaces of information (Chen, Shtykh and
Jin, 2009). RS recommend subjects or items of
interest to users based on information that may be
categorized in two main classes, the first being
gathered from the users explicitly through
interactive collection means (e.g. dialoguing or
conversational modules) and the second being
implied using various modeling and prediction
techniques of user behaviour (Baraglia and Silvestri,
2007).
While there is a plethora of research works
dealing with RS development the majority focuses
on ways to implement the recommendation engine
based purely on technical development factors.
Nevertheless, a RS may be greatly enhanced if
certain factors falling into a non-technical sphere are
taken into account prior and during its creation.
These factors mostly describe human, social,
cultural and organizational (HSCO) aspects which
may complete the picture for understanding the
behavior, as well as the expectations of users
working on the Web, thus contributing to better
modeling certain behavioral characteristics and
ultimately leading to finer information
recommendations. In this context we will view RS
as a specific-purpose (or part of a) Web application
and we will provide a methodology for identifying
and analyzing certain HSCO factors which will be
incorporated in the RS thus enhancing its efficacy.
The methodology utilizes the SpiderWeb model
proposed in (Andreou, Mavromoustakos and
Schizas, 2002), which records critical factors that
must be incorporated as functional or non-functional
features in the RS under development.
2 THE SPIDERWEB MODEL AND
THE RS INFORMATION
GATHERING METHODOLOGY
The SpiderWeb model aims at visualizing and
classifying valuable requirement components for the
211
S. Andreou A. and Mavromoustakos S..
ENHANCING RECOMMENDER SYSTEMS DEVELOPMENT WITH HUMAN, SOCIAL, CULTURAL AND ORGANIZATIONAL FACTORS.
DOI: 10.5220/0003553202110214
In Proceedings of the 13th International Conference on Enterprise Information Systems (ICEIS-2011), pages 211-214
ISBN: 978-989-8425-56-0
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
better identification of critical factors that will lead
to the development of successful software systems
running on the Web. In our case, success is assumed
when the RS is able to provide information
suggestions that are very close to the user interests,
characteristics and expectations at the time of
accessing particular portions of Web information.
The model categorizes system requirements in three
main axons: The Country Characteristics, the User
Requirements, and the Application Domain axon
(fig. 1).
Figure 1: The SpiderWeb Model.
Each axon includes certain components, which are
directly connected and interrelated. The SpiderWeb
axons are also interdependent, allowing the sharing
of same, similar, or different characteristics among
each other.
The analysis of the axon components of the
SpiderWeb model presented in the previous part
aimed primarily at providing the basic key concepts
for collecting proper system requirements. These
concepts are to be used as guidelines for gathering
critical information that may affect the functional
and non-functional behavior of the RS under
development. A form of small-scale ethnography
analysis is conducted for collecting and analyzing
information for the three axons described before.
Our method includes focus questions produced
in the form of questionnaires. These questions are
distributed among the targeted group or are used as
part of the interviewing process, and the answers are
recorded, analyzed and evaluated.
The SpiderWeb methodology is integrated with
the WebE process (Pressman, 2000), the latter being
used for the development of Web applications.
As previously mentioned, the SpiderWeb model
is utilized to guide the creation of the
recommendation engine so as to provide the right
recommendations. The engine employs both offline
and online data gathering and processing procedures
as follows:
Offline Operation Procedures
During the Analysis Phase the following offline
operation procedures are used:
Step 1 – Administer focus questions to groups of
users and collect HSCO information on country
characteristics, user requirements and the RS
itself.
Step 2 – Estimate the preferences’ priorities (e.g.
purchasing decisions or searching patterns)
according to the user groups and the type of the
application
Step 3 – Compute similarity measures between
user groups
Step 4 – Derive the possibility measures
Online Operation Procedures
Step 1 – Set up basic parameters by establishing
an initial dialogue with users to collect HSCO
information (e.g. age, gender)
Step 2 – Record on-click and identify users’
interests and preferences patterns
Step 3 – Analyze search and browse patterns,
build categories
Step 4 – Match related content with categories
Step 5 – Provide recommendations
First, historical data are selected and added into
datasets. For every search, if frequently occurring
patterns (classifiers) are found then those of good
quality are used for recommendations. When new
information is requested, the system identifies the
corresponding class labels using multiple classifiers.
Finally, the performance of the RS is evaluated by
investigating the accuracy of the recommendations
offered.
Typically, a client is accessing the RS through
her/his Web browser where she/he can search and
retrieve information (fig. 2). The Web server will
receive her/his request and information and then
process the data. The requirements pre-processing
subsystem will receive the request and search the
semantic rule database and if there is a relationship
then there is a match to the original requirements.
Offline classifiers are built and stored in the
classifier rule database for new classifications. For
every new customer requirements, the server will
find all the classifiers that meet the conditions of
these requirements by searching the classifier rule
database and the new recommendations are
provided.
Country
Characteristics
User
Re
q
uirements
Application
Domain
1 2 . . .
c
2
1
1
2
r
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
212
Figure 2: The 3-Tier Architecture of our RS.
3 RS CASE STUDY
We have developed an application to provide
information on treatments and health care providers
around the world based on patients’ requests for
medical tourism. Finally, we enhanced the system by
adding a RS to provide more options to patients on
health care providers as well as treatments.
RS are typically either homogeneous (i.e.,
content-based filtering) or heterogeneous (i.e.,
collaborative filtering) for product recommendations
(Yuan and Cheng, 2004; Schafer, Konstan and
Riedl, 1999). The Content-Based Filtering (CBF)
approach recommends products to target customers
according to the preferences of their neighbors while
the Collaborative Filtering recommends products to
target customers based on their past preferences.
We have developed a hybrid RS based on CBF and
CF approaches. Therefore, we combined both
collaborative and content association rules to form
hybrid association rules.
Focusing on the HSCO factors we have isolated
three such factors for demonstration purposes,
namely Gender, Age and Country. A user is
searching the site on the treatment and destination
he/she is interested e.g. Cosmetic Surgery in Cyprus
or browsing the site for the treatment he/she is
interested (fig. 3). The user may select one of the
available providers to check his profile information
such as, description, facilities, medical team, etc.
The RS system provides recommendations for more
available related treatments (on cosmetic surgery)
offered by the provider. In addition, he/she will
receive recommendations on more available
providers that offer the treatment requested, in this
case cheek implants.
If we now take into account the aforementioned
three HSCO factors then recommendations become
more accurate and closer to the user (hidden)
preferences. According to user’s profile the system
provides some recommendations that are directed
towards that specific gender like Masks, Lipsticks,
Makeup etc (fig. 4a). According to Age, a simple
rule inserted into the system as a result of the pre-
processing stage the system suggests certain types of
cosmetic surgery or/and treatments that are highly
popular among a certain range of age, like face
lifting, botox, breast augmentation or lift, etc. (fig.
4b).
Figure 3: Browsing Cosmetic Surgery Treatments.
Country characteristics may also lead to
suggesting other sources of information, like site
seeing in the specific country, monuments etc (fig.
4c). The recommendations are also prioritized based
on the city of his/her choice and neighboring cities
within the country of origin based on distance
proximity.
The system was validated using a population of
75 people both male and female persons, with
average age of 43 years old.
A questionnaire was distributed to the population
to evaluate the effectiveness and efficiency of the
RS. The responses of the users revealed the correct
nature of the recommendations which were made
based on the HSCO factors gathered via the
SpiderWeb. Female subjects of our population were
Client
Web browser
Pre-processing
Classification
Tier 1 Tier 2 Tier 3
DB Rules
Web server
ENHANCING RECOMMENDER SYSTEMS DEVELOPMENT WITH HUMAN, SOCIAL, CULTURAL AND
ORGANIZATIONAL FACTORS
213
Figure 4: Recommendations of information based on (a) Gender, (b) Age, (c) Country.
excited to read the recommendations on cosmetics,
fashion and the alternative treatments (surgical)
suggested by the system. The male subjects did not
receive any such recommendations and were
satisfied they did not have to go through
information of no practical benefit to them. In
addition, ratings of the recommendations ranged
between good and excellent, with people finding
very helpful the suggestions of the system,
understandable and clear. It is worth noting that 5
out of our 35 female persons were reluctant to give
their age (and did not actually provide this data till
the end of the session) as they felt it was a
sensitive personal information. Fortunately, 4 of
our remaining female persons that actually inserted
their age to the system were over 40 and thus they
received the information regarding additional
treatments based on their age range, something
which was found at first amusing but also quite
interesting as these specific persons had limited
knowledge on the suggested therapies. Overall, 72
people replied that they prefer to use a Web
application with recommendations.
4 CONCLUSIONS
In this paper we have proposed a method based on
the SpiderWeb model for identifying significant
human, social, cultural, and organizational
(HSCO) requirements for enhancing the
development of RS.
We have developed a small pilot application
with which the efficacy and efficiency of a RS that
takes into consideration HSCO factors was
demonstrated. A short scale user validation
suggested that indeed when taking such factors
into consideration for producing recommendations
satisfaction between users is greatly enhanced as
the time to search for and locate useful information
is minimized and often the suggested information
broadens the views of users on certain subjects.
Future research will concentrate on the use of
Fuzzy Cognitive Maps (FCM) so as to approach
user behaviour from a quantitative point of view.
REFERENCES
Andreou, A. S., Mavromoustakos, S. M., Schizas, C. N.
2002. Towards the Identification of Human, Social
and Organizational Requirements for Successful E-
Commerce Systems Development. In Proceedings of
the International Conference on Financial
Engineering, E-Commerce and Supply Chain,
Athens, Greece, pp. 24-30.
Baraglia, R., Silvestri, F., 2007. Dynamic
Personalization of Web sites Without User
Intervention. Communications of the ACM, Vol.50,
No.2, pp.63-67.
Jian Chen, J., Shtykh, R., Jin, O. 2009. A Web
Recommender System Based on Dynamic Sampling
of User Information Access Behaviors. 9th IEEE
International Conference on Computer and
Information, Technology, Xiamen, China,, pp. 172-
177.
Pressman, R. S., 2000. Software Engineering: A
Practitioner’s Approach. London: Mcgraw-Hill.
Yuan, S. T., Cheng, C., 2004. Ontology-Based
Personalized Couple Clustering For Heterogeneous
Product Recommendation. In Mobile Marketing.
Expert Systems With Applications, 26, pp.461–476.
Schafer, J. B., Konstan, J. A., Riedl, J., 1999.
Recommender Systems in E-Commerce. In
Proceedings of the 1
st
ACM EC, pp.158–166.
(a) (b)
(c)
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
214