A RECOMMENDATION BASED FRAMEWORK FOR ONLINE
PRODUCT CONFIGURATION
Nikos Karacapilidis
IMIS Lab, MEAD, University of Patras, 26500 Rio Patras, Greece
Thomas Leckner
Institut für Informatik, Technische Universität München, Boltzmannstr. 3, D-85748 Garching, Germany
Keywords: Product Configuration, Recommender Systems, Personalization, Similarity Measures, Product Modelling
Abstract: Adopting a mass customization strategy, enterprises often enable customers to specify their individual
product wishes b
y using web based configurator tools. With such tools, customers can interactively and
virtually create their own instance of a product. However, customers are not usually supported in a
comprehensive way during the configuration process, thus facing problems such as complexity, uncertainty,
and lack of knowledge. To address the above issue, this paper presents a framework that aids customers in
selecting and specifying individualized products by exploiting recommendations. Having first focused on
the characteristics of configurator tools and the principles of model-based configuration, we then introduce
the concept of masks for product models. The main contribution of this paper is the proposal of an
integrated approach for supporting model-based product configurator tools by similarity-based
recommendations. Our approach in providing recommendations has been based on the widely accepted
theory of Fuzzy Sets and its associated concept of similarity measures, while recommendations provided are
based on the processes of stereotype definitions and dynamic customer clustering.
1 INTRODUCTION
Mass customization, as a business strategy, aims at aiding
companies to react to the growing individualization of
demand, by giving them a more customer-centric role
(Piller, 2001; Pine, 1993). At the same time, it aims at
providing individualized products at a price which is close
to that of standardized products. The adoption of such a
strategy is admittedly associated with the need for major
changes in various perspectives, such as product design
and manufacturing, technology and innovation
management, marketing, logistics, and information
management.
On the other hand, there are also problems in
supporting the customer to
express his wish for an
individual product. The customer should easily become
aware of a product’s “degrees of freedom”, while he needs
tools that enable him to translate his product wish to an
instance of a predefined product family. Compared to
custom-made products, companies cannot afford to offer
professional consultants to help the customer in the above
process. To address the above issues, there is a trend for
developing and deploying online configurator tools that
support customers to go through an interactive and virtual
individualization of a product by using an internet
browser. However, configurator tools demonstrate a series
of shortcomings such as confusion, frustration and
uncertainty that often push customers towards the
dropping out of the configuration process.
To overcome such problems and efficiently support a
customer during
the process of configuring a product, the
approach discussed in this paper builds on the integration
of configuration and recommendation features.
Recommendations are based on predefined stereotypes
that best match the individual customer’s profile. The
remainder of the paper is structured as follows: Section 2
comments in detail on the functionality and the limitations
of configurator tools. The representation of a product’s
model and the underlying configuration and
recommendation issues are described in Section 3. The
proposed framework and its associated processes are then
comprehensively discussed in Section 4. Finally, Section 5
outlines final remarks and future work directions.
303
Karacapilidis N. and Leckner T. (2004).
A RECOMMENDATION BASED FRAMEWORK FOR ONLINE PRODUCT CONFIGURATION.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 303-308
DOI: 10.5220/0002643203030308
Copyright
c
SciTePress
2 CONFIGURATOR TOOLS
2.1 Basic functionality
Online configurator tools provide the essential means for
supporting mass customization, by enabling a customer to
virtually assemble a product according to his individual
needs and preferences (Sabin and Weigel, 1998). Well-
known examples of such tools can be found in the
homepages of big automakers and computer vendors.
Usually, a configurator tool is built around a specific form
of product model. Therefore, one of the most fundamental
functionalities of a configurator tool is to manage and
represent the underlying product and configuration models
(Tiihonen and Soininen, 1997). The product model
represents the product’s physical and logical structure,
which usually has been predefined by the manufacturer.
On the other hand, the configuration model represents the
customer’s current instance of the product model, which is
shaped upon the customer’s selections and restrictions.
Another basic functionality of a configurator tool is to
provide customers with an overview of the available
“degrees of freedom” and, more important, to enable
customers to manually configure and manipulate them.
Moreover, configurator tools often integrate mechanisms
for checking the correctness of a configuration model.
These mechanisms exploit methods and algorithms
originally coming from the Artificial Intelligence
discipline to deal with problems such as constraint
checking and constraint satisfaction (Felfernig et al.,
2001).
2.2 Enhanced functionality
In addition to the above fundamental functionalities, there
are tools that demonstrate some more advanced features.
More specifically, a configurator tool may provide access
to a database of previously configured products and
components via different catalogue systems. For instance,
such a database may include participatory catalogues
(Schubert, 2000) that are enriched with ratings and/or
comments of other customers and can be filtered by
special orders (e.g., according to the name of the
customer, the average rating of certain groups etc.). In
such a way, the customer gets access to the collective
knowledge of the community of customers and can take
into account their opinions and experiences for the
individual decision making (Leckner, 2003).
It might be also possible that the configurator tool
assists the customer during the process of configuration on
the basis of the product model. This means that the
customer makes the main decisions, while the system
propagates their consequences and somehow “explains”
the product model to the customer (Inakoshi et al., 2001).
Further automation of the configuration process is also
possible by such approaches, where the customer makes
only some basic decisions and the system configures the
rest of the product automatically (Ardissono et al., 2001).
Configurator tools may also provide recommendations
to the customers in an active manner. To give an example,
this can be performed through personalised defaults for the
product’s degrees of freedom or through restricted product
models (see Section 3.3). Different approaches for the
creation of such recommendations have been proposed in
the literature, which address the issue independently or in
a hybrid mode, combining more than one of such methods
(Renneberg and Borghoff, 2003). Personalized
recommendations are always based on further information
about the customer, which is stored in the customer’s user
profile.
2.3 Limitations
Configurator tools allow the specification of form, fit,
function and modalities of a product (Leckner, 2003). In
the ideal case, the customer can enter directly the above
specification into the company’s information system.
However, configurator tools demonstrate some limitations
and shortcomings. First of all, the manual configuration of
a product usually takes a lot of time and effort from the
customer, especially if the product is complex. This is the
case with products characterized by multiple degrees of
freedom, which is a very likely scenario in the
contemporary marketplaces and mass customization
MASK1 MASK2
(a) (b) (c)
Figure 1: Masks for alternative and optional components
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
304
initiatives.
Another problem with configurator tools is that even if
the customer is willing to spend enough time, he often
lacks the know-how and experience in using the tool and
properly configure the product according to his individual
preferences and needs. Even worse, assuming that the
customer has enough time, know-how and experience, he
often does not exactly know what he wants. This is a
general problem with such tools, since the customer is
about to configure something complete new, which he
cannot see, feel or test until he buys it. For the above
reasons, the process of virtually configuring a product
often leads to confusion, frustration, uncertainty and,
consequently to the abandonment of the configuration
process (Huffman and Kahn, 1998; Piller et al., 2003).
To overcome these problems, customers should be
supported in a more efficient and effective way during the
process of configuring a product. An approach to
overcome the customer’s confusion and uncertainty builds
around the concept of virtual communities, where
customers support each other during the overall process
(Rheingold, 1998; Leckner, 2003). This approach is
further motivated by the fact that individual decisions
often depend on decisions made by others (Wind et al.,
2001). Another approach to be exploited concerns
provision for recommendations that help the customer
shape his decision in an automatic way. In some respects,
such recommender systems also exploit Artificial
Intelligence techniques. Section 4 is particularly devoted
to such recommendations, the aim being to overcome the
customer’s uncertainty, confusion and frustration.
3 PRODUCT MODEL
3.1 Representation of the product
The backbone of every configurator tool is the underlying
product model, which represents the product’s physical
and logical structure. Additionally, such a model defines
the associated degrees of freedom, which are actually the
product’s elements that can be directly modified by the
customer. Representative examples are the attributes of a
product as well as its alternative and optional components.
Every degree of freedom can have a range of valid values
and a default value. Moreover, certain restrictions and
interdependencies between different degrees of freedom
are possible (Männistö et al., 2001). In the ideal case, the
product model contains all the product-related
“knowledge” about the product, while the configurator
tool can provide this knowledge in an appropriate way to
the customer (Tiihonen et al., 1998).
The product model is initially defined by professional
product designers of the manufacturer. We assume that
every physical product consists of a set of components, the
connected structure of which can be described by a
component-tree. Each component can in turn consist of a
set of components and/or a set of attributes. Attributes can
be based on various data types. For example, attributes of
numeric interval type are defined by an upper boundary, a
lower boundary and a default value. The configuration
model derives from the product model by incorporating
the customer’s selections regarding the associated degrees
of freedom.
Fi
g
ure 2: The data flow dia
g
ra
m
of our a
pp
roach.
A RECOMMENDATION BASED FRAMEWORK FOR ONLINE PRODUCT CONFIGURATION
305
3.2 Restricted product model
Although the idea of model-based product development is
not new (Anderl and Trippner 2000; Felfernig et al.,
2001), special requirements have to be taken into account
when enabling the customer to configure his individual
product. On the one hand, the product model should not be
too complex, so that the customer can understand and
manipulate it. On the other hand, the product model must
not be too simple; otherwise, the customer has not enough
possibilities to express his individual product wishes and
preferences. In the ideal case, every customer will get a
personalized version of the product model, which is
restricted in accordance to the customer’s individual needs
and interests.
This idea leads to the concept of masks for product
models, which is comprehensively described in (Leckner
and Lacher, 2003). Figure 1(a) depicts a specific
component-based product model in its entirety, whereas
Figures 1(b) and 1(c) two valid masks of it (
MASK1 and
MASK2, respectively). As shown, we follow a tree-like
representation of a product model. The degrees of freedom
of
MASK1 and MASK2 are obviously smaller than that of in
the original product model. We should note here that each
component of the model is associated with a set of
attributes, which may in turn impose additional degrees of
freedom. Finally, one can easily understand that some
masks correspond to a completed product configuration
where no alternative or optional components exist (e.g.,
MASK1), while others need further manipulation by the
customer (e.g.,
MASK2).
Another easily conceivable example for the usage of
masks on product models can be given through a product’s
attribute of numeric interval type. Let the possible values
for the power of an engine be constrained to the interval
[40...300 kW] in the product model. However, restrictions
imposed for a specific user may result in possible values
within the interval [110...200 kW]. The restrictions for
another user may result in a different interval, say
[80…130 kW]. In addition, the default value of this
attribute can be also personalized according to the specific
customer.
Such a restricted product model can be seen as a type
of recommendation. Another type of recommendation is to
personalize the selected values of each degree of freedom
in accordance to the customer’s user profile. Restricted
product models together with personalized selected values
also can be predefined by the manufacturer to satisfy
certain stereotypes of customers with adequate product
models. Concepts for assigning customers to a certain
stereotype and ideas about defining such stereotypes
dynamically are discussed in the following section.
4 THE OVERALL FRAMEWORK
Our approach (see Figure 2) maintains a detailed profile
for each customer. Such profiles contain information about
the customer’s basic and demographic attributes, general
interests and life style, specific product interests and
buying history. Some pieces of such information are
gathered when the customer uses the proposed system for
the first time, through a specially constructed
questionnaire, while others through the customer’s
interaction with it. Moreover, based on predefined rules,
the company maintains a set of stereotypes, which are
basically based on attributes related to the profession,
general interests and life style of the customers.
Stereotypes are associated with the product models
stored in the company’s database and are used to restrict
their degrees of freedom (they are actually associated with
predefined product model masks). An example illustrating
the definition of two such stereotypes, namely “engineer”
and “lawyer”, is given below.
public String stereotype(UserProfile customer) {
String job=customer.basic.business.jobtitle;
// definition of stereotype “engineer”
if ((job==“engineer”) or (job==“mechanist”)
or (job==“electrician”) or (job==“designer”)
or (job==“natural scientist”)
or (customer.interest.computers>70))
return “engineer”;
// definition of stereotype “lawyer”
if ((job==“lawyer”) or (job==”magistrate”)
or (job==“judge”) or (job==“tax inspector”)
or (customer.interest.computers<30))
return “lawyer”;
...
}
// assign stereotype to product model
switch (stereotype(customer)) {
case “engineer”: ProductModel=MASK1; break;
case “lawyer”: ProductModel=MASK2; break2;
...
}
As shown, our approach uses a set of rules that are
initiated upon the customer’s profile information
(information about one’s profession and interest in
computers is only used in this example). The second part
of the above example corresponds to the association of
masks (like those shown in Figure 1) to the two previously
defined stereotypes.
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
306
4.1 Stereotype-based
recommendations
To aid customers better configure a product, our approach
exploits the concept of fuzzy similarity measures to decide
how close a customer is to a predefined stereotype. Based
on the results of this procedure, the system recommends to
the customer the product mask(s) that is (are) associated to
the stereotype the customer is more close.
More specifically, each stereotype defined by the
company corresponds to a fuzzy set A that is structured
according to the scores assigned for each individual
customer’s attribute considered. To more realistically
decide about the attributes to be considered and the scores
to be assigned, the company may go through appropriate
web mining and knowledge construction processes (Cho et
al., 2002; Nahm and Mooney, 2002). While of much
importance, such processes are out of the scope of this
paper. Instead, we concentrate on the process of
associating a predefined stereotype to a specific customer.
Each time a customer uses the system, our framework
extracts information from his profile to construct a fuzzy
set B, which also encapsulates the scores assigned to the
attributes under consideration (i.e., the attributes used for
the definition of stereotypes). In fact, these scores express
the magnitude in which a customer is interested in or
attracted by the set of attributes F
i
that characterize each
product O
j
.
To give an example, the choices offered to the user
can be in the set {minimal interest, less interest, neutral,
much interest, extensive interest}, where each choice is
associated to a value in the interval [0.2 … 1.0]
(alternative ratings, following different granularity levels,
may be also applied). The above fuzzy set is associated
with the specific user and is stored in the system’s
database.
The next step concerns the comparison of a specific
customer’s set B with each of the predefined, stereotype-
based fuzzy sets A. To do this, our approach uses the
similarity measure Q
A,B
that is based on the difference of
grades of membership and the volume of the two fuzzy
sets. The Q
A,B
similarity measure was selected among
various fuzzy sets similarity measures (Wang, 1997), after
evaluating their properties. Q
A,B
was the only one which
was a proximity measure (a normalization of the
attributes’ ranges and/or values allowed is required when
applying this measure). In fuzzy sets theory, a similarity
measure is called a proximity measure when it stands:
Q
A, B
= Q
A^, B^
where A^ and B^ are the supplements of A
and B, respectively. Using this proximity measure, we can
efficiently consider the influence of both high and low
similarity. It is:
niBiAabsQ
n
i
BA
/}))]()((1[{
1
,
=
=
(1)
where A(i) is the score of attribute i stored in the
database (that is, the one that corresponds to the
stereotype), B(i) the score of the same attribute that is
given by the user, and n the total number of attributes for a
stereotype. The results of this process are temporarily
stored in a table containing the score (similarity) of each
stored stereotype against the user’s preferences. In the
sequel, the best N results and their associated scores are
retrieved (number N is defined by the user).
In other words, our approach provides the N most
similar stereotypes to the description given by the user.
Our framework may also provide another round of
recommendations, by classifying the N retrieved best
stereotypes according to each individual attribute. As a
result, the customer is informed about which of the N
stereotypes performs better according to each attribute and
may get motivated for further contemplation. It should be
noted here that at this stage the customer may have still to
decide about the alternative or optional parts of a product
mask.
4.2 Customer clustering-based
recommendations
To further aid customers configuring a product, our
approach also provides them with the ability to retrieve
recommendations based on data extracted through the
process of customers clustering (see Figure 2). Upon the
users’ wish, the system is able to further help them in
making up their mind, by retrieving and providing
information regarding products already bought by
customers with similar profiles. The results provided here
are restricted to products of the same category. As in
stereotype-based recommendations, the similarity measure
described in Equation (1) is applied here to retrieve similar
customers.
5 CONCLUSION
The main contribution of this paper is the proposal of an
integrated framework for supporting model-based product
configurator tools by similarity-based recommendations.
We have first highlighted issues involved in the process of
creating personalized recommendations to support
customers virtually specifying products with a
configurator tool. We then discussed basic functionalities
and shortcoming of existing configurator tools, and we
introduced a set of product modelling aspects. Our
approach in providing recommendations has been based
on the widely accepted theory of Fuzzy Sets and its
associated concept of similarity measures, while
recommendations provided are based on the processes of
stereotype definitions and dynamic customer clustering.
Future work directions concern the complete
integration of the proposed recommendation methods in
our already implemented product configuration system. In
A RECOMMENDATION BASED FRAMEWORK FOR ONLINE PRODUCT CONFIGURATION
307
addition, we intend to incorporate the concept of fuzzy
similarity measures into the filtering pipeline (for more
details on this issue, see (Stegmann et al., 2003)). Other
important issues to be explored in the future concern the
development and deployment of innovative methods for
customer profile acquisition (based on methods of natural
language processing) and the enhancement of model-based
configurator tools by community functionalities (Leckner,
2003). The ultimate aim of our overall research effort is to
more efficiently support the customer during the process
of configuring a product, thus enabling him to make
rational choices that - as much as possible - express his
wishes and interests.
ACKNOWLEDGMENTS
Parts of this work belong to a larger research effort
towards local production of individualized products,
funded by the German Research Foundation (Deutsche
Forschungsgemeinschaft, DFG) in the program SFB582 –
“Marktnahe Produktion individualisierter Güter”. See
http://www.sfb582.de/ for more information about the
research effort.
REFERENCES
Anderl, R. and Trippner, D. (2000). STEP – STandard for
the Exchange of Product model data, Teubner,
Stuttgart, Germany.
Ardissono L., Felfernig A., Friedrich G., Jannach D.,
Zanker M. and Schäfer R. (2001). Intelligent
interfaces for distributed web-based product and
service configuration, In: Proc. of the 1st Asia-Pacific
Conference on Web Intelligence (WI-2001), Japan
2001.
Cho, Y., Kim J. and Kim, S. (2002). A personalized
recommender system based on web usage mining and
decision tree induction. Expert Systems with
Applications, Vol. 23, pp. 329-342.
Felfernig A., Friedrich G. and Jannach D. (2001).
Conceptual modeling for configuration of mass
customizable products, Artificial Intelligence in
Engineering, Vol. 15 (2001), 2, pp. 165-176.
Huffman C. and Kahn B. (1998). Variety for sale: mass
customization or mass confusion. Journal of Retailing,
Vol. 74, No. 4, pp. 491-513.
Inakoshi H., Okamoto S., Ohta Y. and Yugami N. (2001).
Effective Decision Support for Product Configuration
by Using CBR, Fujitsu Labs, Japan 2001.
Leckner T. (2003). Support for online configurator tools
by customer communities, Proc. 2nd Intl. World
Congress on Mass Customization and Personalization
(MCPC’03), Munich, Germany.
Leckner T. and Lacher M. (2003). Simplifying
configuration through customer oriented product
models, Proc. of the 14th International Conference on
Engineering Design (ICED 2003), Stockholm,
Sweden.
Männistö T., Soininen T. and Sulonen R. (2001).
Modeling Configurable Products and Software
Product Families, Software Configuration Workshop
(SCM-10) of ICSE01, Toronto, Canada.
Nahm, U. and Mooney, R. (2002). Text Mining with
Information Extraction. In Proc. of Spring Symposium
on Mining Answers from Texts and Knowledge Bases,
Stanford, CA, pp. 60-68.
Piller F. (2001). Mass Customization. Ein wettbewerbs-
strategisches Konzept im Informationszeitalter,
Wiesbaden: Gabler Deutscher Universitäts-Verlag.
Piller F., Koch M., Möslein K. and Schubert P. (2003).
Managing High Variety: How to Overcome the Mass
Confusion Phenomenon of Customer Co-Design,
EURAM 2003.
Pine J. (1993). Mass Customization: The New Frontier in
Business Competition, Boston: Harvard Business
School Press.
Renneberg V., Borghoff U. (2003): Recommendations in
Product Personalization: An extended Example of
Pipelined Filter Combination, Proc. 2nd Intl. World
Congress on Mass Customization and Personalization
(MCPC’03), Munich, Germany.
Rheingold H. (1998). The Virtual Community, free in the
Web at:
http://www.rheingold.com/vc/book/
Sabin D. and Weigel R. (1998). Product Configuration
Frameworks – A Survey, IEEE Intelligent Systems &
their applications 13(4), pp. 42-49.
Schubert P. (2000). The Participatory Electronic Product
Catalog: Supporting Customer Collaboration in E-
Commerce Applications, In: Electronic Markets
Journal, Vol.10, No.4
Stegmann R., Koch M., Lacher M., Leckner T. and
Renneberg V. (2003). Generating Personalized
Recommendations in a Model-Based Product
Configurator System, Proc. of Workshop on
Configuration held in conjunction with 18th
International Joint Conference on Artificial
Intelligence (IJCAI-03), Acapulco, Mexico.
Tiihonen J. and Soininen T. (1997). Product
Configurators – Information Systems Support for
Configurable Products, in: Increasing Sales
Productivity through the Use of Information
Technology during the Sales Visit – A Survey of the
European Market, Hewson Consulting Group, 1997.
Tiihonen J., Lehtonen T., Soininen T., Pulkkinen A.,
Sulonen R. and Ritahuhta A. (1998). Modeling
configurable product families, 4th WDK Workshop on
Product Structuring, Delft University of Technology,
Delft, Netherlands.
Wang, W.J. (1997). New similarity measures on fuzzy sets
and on elements. Fuzzy Sets and Systems, Vol. 85, pp.
305-309.
Wind Y., Mahajan V. and Gunther E. (2002).
Convergence Marketing – Strategies for Reaching the
New Hybrid Consumer, Prentice Hall, USA.
ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING
308