Koba4MS: Knowledge-based Recommenders
for Marketing and Sales
Alexander Felfernig
Computer Science and Manufacturing, University Klagenfurt,
Universitätsstrasse 65-67, A-9020 Klagenfurt, Austria,
Abstract. Due to the increasing size and complexity of products offered by on-
line stores and electronic marketplaces, the identification of solutions fitting to
the wishes and needs of a customer is a challenging task. Customers can dif-
fer greatly in their expertise and level of knowledge w.r.t. the product domain
which requires sales assistance systems allowing personalized dialogs, explana-
tions and repair proposals in the case of inconsistent requirements. In this context,
knowledge-based recommenders allow a flexible mapping of product, marketing
and sales knowledge to the formal representation of a knowledge base. This pa-
per presents the knowledge-based recommender environment Koba4MS which
assists customers and sales representatives in the identification of appropriate so-
lutions. Based on application examples from the domain of financial services,
basic Koba4MS technologies are presented which support the effective imple-
mentation of customer-oriented sales dialogs.
1 Introduction
Buying complex products (e.g. financial services, computers, etc.) is still a challenging
task since many organizations offer simple query interfaces under the assumption that
customers know the technical details of the offered set of products [1]. Recommender
technologies [1,2,3,4,5,6,7,8,9] improve this situation by providing solution alternatives
for the customer which are automatically derived from a set of customer requirements.
There are three basic approaches to the implementation of recommender applications.
Collaborative Filtering [4,6,7] is based on the concept of storing preferences of a large
set of customers. Based on the assumption that human preferences are correlated, rec-
ommendations given to a customer are derived from preferences of a group of customers
with similar interests, i.e. no deep knowledge about product properties is needed. Sim-
ilarly, using Content-based Filtering [3,5], products are described by a set of keywords
(categories) which are stored in a customer profile in the case that a customer buys a
certain product. The next time, the customer enters the system, the stored preferences
are used for identifying additional products which are assigned to similar categories.
Additionally, there exist a number of approaches combining these basic approaches
in order to gain an improved quality of the resulting solutions (see e.g. [10]). Finally,
Knowledge-based Recommender applications (advisors) [1,2] exploit deep knowledge
about the product domain in order to determine solutions exactly fitting to the wishes
and needs of the customer. When selling complex products such as financial services, a
Felfernig A. (2005).
Koba4MS: Knowledge-based Recommenders for Marketing and Sales.
In Proceedings of the 1st International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, pages 164-174
DOI: 10.5220/0001422601640174
Copyright
c
SciTePress
customer’s taste is not of primary concern - primarily solutions and explanations must
be correct in every case (e.g., due to legal regulations), i.e., must correspond with the
requirements articulated by the customer (see Section 3.1 for the repair of inconsistent
customer requirements). This requirement can only be met by explicitly representing
product, marketing, and sales knowledge [11,12], i.e. Knowledge-based Recommender
applications (advisors) are the natural choice in this context.
Fig.1. Overall architecture.
In the following we give an overview of the major technologies implemented within
the Koba4MS
1
environment, a domain-independent tool designed for the development
of knowledge-based advisors. The major difference between Koba4MS and other know-
ledge–based recommender systems [2] is the inclusion of model-based diagnosis [13,14]
and personalization techniques [1] which improve the effectiveness of advisor devel-
opment as well as the interaction with the advisor. For example, a graphical devel-
opment and test environment makes the implementation of advisors feasible for non-
programmers, furthermore intelligent diagnosis and repair techniques actively support
customers in situations where no solution could be found. Koba4MS can be applied
in the following scenarios. Firstly, similar to traditional sales channels, improved sales
assistance by improving the accessibility of a product assortment on a companies Web
page generates added value for customers. On the one hand advisors allow an intu-
itive access to complex products for customers, on the other hand sales representatives
are relieved from routine advisory jobs. Secondly, sales representatives interact with
advisors when talking with the customer, where guided dialogs provide questions and
explanations focusing on the customers wishes and needs. Throughout the paper we
provide real-world examples from the financial services domain which is our leading
application domain. Financial service advisory is a knowledge-intensive task which in
many cases overwhelms customers as well as sales representatives. Therefore financial
service providers ask for tools providing an intuitive access to their product assortment.
2
The remainder of the paper is organized as follows. In Section 2 we present the
Koba4MS recommender environment. In Section 3 we present examples for the us-
age of AI technologies which allow the implementation of knowledge-based advisors
1
Knowledge-based Advisors for Marketing and Sales is a project funded by the Austrian Re-
search Fund (agreement number FFF-808479).
2
Products of financial service providers cover different areas of interest such as investment
decisions, financing, pension, life insurance, etc.
165
(constraint satisfaction, model-based diagnosis and test, personalization, knowledge ac-
quisition). Finally, Section 4 presents experiences from commercial advisor projects.
2 Koba4MS Environment
The Koba4MS environment (see Figure 1) provides comprehensive assistance for cus-
tomers and sales representatives by supporting guided and personalized dialogs allow-
ing an intuitive access to an assortment of complex products. In this context Koba4MS
can be used for the following purposes.
formalization of product, marketing and sales knowledge by non-programmers.
testing/debugging knowledge bases in order to identify faulty constraint definitions.
checking customer requirements for consistency and (in the case of inconsistencies)
supporting a corresponding error handling.
matching customer requirements to product properties, i.e. calculating a solution.
diagnosing and repairing a set of inconsistent customer requirements, i.e. proposing
minimal changes which allow the retrieval of a solution.
explaining solutions in order to increase the confidence of the customer.
Koba4MS technologies are used in application domains such as financial services, dig-
ital cameras, cigars, computers, services in public administration etc. In the financial
services domain Koba4MS technologies are applicable for the following reasons.
Solutions must be objective, correct and explainable which makes approaches such
as Collaborative Filtering or Content-based Filtering not the best choice.
Typically, financial service providers want to develop advisors autonomously, i.e.
knowledge representation formalisms are needed which allow the development of
recommender knowledge bases for non-programmers (this is supported by graphi-
cal knowledge acquisition, model-based debugging and testing).
Intelligent explanation, debugging, and repair mechanisms as well as automated
test case generation are using model-based knowledge representations, i.e. deep
knowledge about the application domain must be available (which is not available
in Collaborative Filtering or Content-based Filtering approaches).
Financial services recommendation is a complex task with a large number of con-
straints and possible solutions. In this context, knowledge-based approaches can
significantly reduce efforts related to advisor development and maintenance.
Similar reasons motivate the application of knowledge-based advisors in other applica-
tion domains such as online-selling of computers, digital cameras, etc.
2.1 Overall Architecture
Koba4MS product knowledge bases and process definitions are developed and main-
tained using a Development and Test environment (Koba4MS Designer and Process
Designer). Products are defined within Koba4MS Designer or imported from exter-
nal systems using an XML interface (for details see [1]). In the following advisors
166
are automatically generated and made available for customers (e.g. online-stores, e-
marketplaces, etc.) and sales representatives (e.g. intranet applications or installations
on notebooks of sales representatives), where Koba4MS Server supports the execution
of advisory sessions (Runtime Environment).
2.2 Development & Test Environment
Koba4MS Designer. Koba4MS Designer is a graphical development environment for
knowledge-based recommenders. It is based on Java Web Start which provides a browser-
independent architecture for deploying Java-2 based applications on a client. The con-
cepts implemented in Koba4MS are based on long-term AI research in the area of
knowledge-based configuration and personalization [1,15,13,11,16]. Koba4MS Designer
supports the design of advisors, where the relevant set of product- and customer prop-
erties is identified and transformed into a recommender knowledge base [11,12]. Such
a knowledge base consists of the following parts (see Figure 2).
product properties are structural descriptions of the provided products (e.g. life in-
surances can be characterised by the possible length of life assurance policies, pre-
miums of life assurance policies, links to additional product documentation, etc.).
customer properties are descriptions of possible customer requirements (e.g. within
an investment advisory process the question under the assumption that your invest-
ment of 10.000 EUROS decreases in value, at which value would you sell your
investment? is related to the willingness to take risks).
constraints are restricting the combinations of customer requirements and product
properties, e.g. return rates above 9 percent require the willingness to take risks.
Constraints can be defined on the graphical level as well as on the textual level.
In order to support the analysis of advisors, Koba4MS Designer provides a statistical
analysis component which operates on interaction logs of advisory sessions conducted
by online customers or sales representatives.
Process Designer. A recommender process represents possible navigation paths which
define the way the system adapts its dialog style to the knowledge level and interests
of the customer. Such process definitions are based on a predicate augmented finite
state recognizer (PFSR) [17] (constraints describe transitions between different states
of a recommender process) which represents allowed navigation paths within an ad-
visor (see Figure 2). Transition conditions between states of a recommender process
are evaluated using the Koba4MS constraint engine. Based on a layout template defin-
ition, knowledge bases and process definitions can be automatically (no programming
is needed) translated into an executable advisor (see e.g. Figure 3), where each state of
the process definition corresponds to a Web-page in the generated application.
Testing & Debugging Knowledge Bases. The increasing size and complexity of recom-
mender knowledge bases makes testing a critical task [18] in the context of successfully
deploying and maintaining recommender applications. Process definitions (see e.g. Fig-
ure 2) are the basis for automatically generating test cases. Solutions (results calculated
167
by the knowledge base) for generated test cases are presented to the domain expert who
decides on their validity (Result Validation). Test cases deemed as correct by the do-
main expert are used for regression tests. Test case generation in Koba4MS follows a
path-oriented approach (the test cases for each path are derived from the set of solu-
tions to a corresponding constraint satisfaction problem) which allows a high degree of
coverage [19]. The disposable time for testing is restricted, consequently mechanisms
are provided which reduce the amount of tests without reducing the coverage of the
overall test suite (except for random selections). Typically, domain experts agree with
accepting efforts related to testing since solution quality is of serious concern. We can
calculate a complete set of test cases which includes all possible transitions of a process
definition, but this is only feasible for small and strongly constrained recommendation
tasks. Approaches reducing the number of test cases are presented in [20].
Fig.2. Definition of customer properties, constraints, recommender processes.
2.3 Runtime Environment
Koba4MS Server. The calculation of solutions for a recommendation task is based
on constraint satisfaction problem solving [21]. Customer properties as well as product
properties are represented as constraint variables. A solution for a given recommenda-
tion task (constraint satisfaction problem) is found if all constraints are satisfied. For an
example screenshot of an interface see Figure 3. Note that Koba4MS allows the logging
of advisory sessions which supports the improvement/fine-tuning of the actual know-
ledge base. Interaction logs can be analysed using the statistical analysis component
integrated in Koba4MS Designer.
168
3 Used Technologies
Compared to Knowledge-based Recommender applications [1,2], Collaborative Filter-
ing [4,7] and Content-based Filtering [3,5] do not exploit deep knowledge about the
domain in order to determine solutions fitting to the wishes and needs of the customer.
Using knowledge-based approaches, the relationship between customer requirements
and financial services can be explicitly modelled in an underlying knowledge base [11].
Such model-based representations are the precondition for applying diagnosis and test-
ing techniques.
3.1 Constraint Satisfaction
Search for Solutions. As already mentioned, Koba4MS problem solving is based
on constraint satisfaction problem solving. A Constraint Satisfaction Problem (CSP)
(C,V ,D) [21] is defined by a set V of variables x
i
, a set C of constraints c
j
and a
set D of domains d
i
which defines for each variable the set of possible values. A CSP
is solved if there exists a set of instantiations of the variables x
1
, x
2
, ..., x
n
s.t. all
constraints contained in C are satisfied. If no solution can be found by the search en-
gine, constraints are relaxed starting with constraints with lowest priority. If nothing but
non-relaxable constraints remain and no solution was found, a repair mechanism is acti-
vated. Solutions are presented in order of their usability for the customer (see Section 3,
Utility of Solutions)
3
. In addition to constraints, Koba4MS supports tips, i.e. constraints
representing e.g. cross-selling opportunities which are presented to the customer with-
out interrupting the recommender process. Cross-selling conditions are integrated in the
recommender knowledge as well represented as special types of constraints. A tip is:
long-term investments reduce risks, i.e. allow higher return rates than short-term invest-
ments without taking high risks. An example for a tip representing a cross-selling rule
is a married sole wage earner with two children taking out a loan, is also a candidate
for a risk insurance.
Diagnosis and Repair of Requirements. If the result set is empty, conventional recom-
menders tell the user (customer) that no solution was found, i.e. no clear explanation
for the reasons for such a situation is given. Koba4MS supports the calculation of repair
actions for customer requirements (a minimal set of changes allowing the calculation
of a solution). If Σ = {x
1
= a
1
, x
2
= a
2
, ..., x
n
= a
n
} is a set of customer require-
ments (Σ C has no solution), a repair is a minimal set of changes to Σ (resulting in
Σ’) s.t. Σ C has a solution. The computation of repair actions [13] is based on the
Hitting Set algorithm [14] which exploits minimal conflict sets (minimal sets Π Σ
of variable instantiations triggering an inconsistency with C) provided by the constraint
solver in order to determine minimal diagnoses and corresponding repair actions.
Test Case Generation. Automated test case generation (for more details see Section
2.2) is based on the definition of a constraint satisfaction problem. For this purpose a
(complete) set of possible paths through a recommender process is determined. For each
3
Critiquing of solutions [8] is within the scope of future versions of Koba4MS.
169
path a corresponding CSP is generated and executed - identified solutions represent test
cases, i.e. possible settings of customer requirements.
3.2 Personalization Concepts
Handling of Profiles. Koba4MS includes mechanisms allowing the adaptation of the di-
alog style to the user’s skills and needs [1]. The user interface relies on the management
of a user model that describes capabilities and preferences of individual customers.
Some of these properties are directly provided by the user (e.g. name or personal goals,
or self-estimates such as knowledge about financial services), other properties are de-
rived using personalization rules and scoring mechanisms which relate user answers to
abstract dimensions [1] such as preparedness to take risks or interest in high profits (di-
mensions describing the users interests) and knowledge about funds, etc. (dimensions
describing the users knowledge about the domain). Values of the customer profile are
collected from recommender sessions, missing values are asked the customer if they are
needed in the context of a certain recommendation process (e.g. if the customer has not
specified his/her age up to now then the related question is posed to the customer in the
advisory dialog).
Dialog Style. Customers have different approaches to specify their requirements rang-
ing from the direct specification of product parameters (e.g. a certain savings account
running for 3 years) to a general specification of their personal goals (e.g. financing their
children’s education). An adaptation of the interaction style can significantly contribute
to an improved approximation to the behavior of a human sales expert (an experienced
sales assistant adapts his dialog style to the skill level and interests of customers). De-
pending on answers already provided by a customer, the dialog style can be personal-
ized as follows.
Alternative formulation of questions, e.g. questions posed to expert users can be
differentiated from those posed to customers with less knowledge about the product
domain.
Rule-based formulation of default-answers, e.g. if the goal of the customer is to put
money by for a rainy day the default answer to a question related to the maximum
accepted decrease in value of the investment is no value decrease accepted.
Alternative explanations for constraint violations, e.g. if the customer is a novice, a
very general explanation about changes in the pension law is given, more detailed
information can be included for experts.
Utility of Repair Proposals. If no solution can be found for a given set of customer re-
quirements (although all relaxable constraints have been relaxed), Koba4MS provides a
set of possible (minimal) repair actions which allow the calculation of a solution. Differ-
ent customer properties have an assigned priority which indicates the importance of the
variable for the customer. The lower the priority of the variable the higher the probabil-
ity is that the variable is considered as focus of repair actions, e.g. if the type of returns
on investment (at the end of the investment period, dividend payout) is not important
170
<< Back Next >>Restart
Accept
customer
properties
possible
answers
explanations for
questions
control 
elements
provisional
results
<< Back Restart
repair
proposals
Fig.3. Example user interface.
for a customer, this property is primarily considered as a potential candidate for repair
actions. More formally, the personalization of repair proposals is based on the formula
f(x
1
, x
2
, ..., x
m
) =
P
m
j=1
p(x
j
), where f(x
1
, x
2
, ..., x
m
) represents the utility of repair
actions related to the variables x
1
, x
2
, ..., x
m
and p(x
j
) denotes the customer-specific
priority of variable x
j
. Customer-specific priorities can be either defined statically or
by a customer within the scope of an advisory session.
Utility of Solutions. A solution for a recommendation task is a set (portfolio) of finan-
cial services. The order of solutions should strictly correspond to the degree a solution
contributes to the wishes of a customer. Koba4MS supports multi-attribute object rat-
ing [1], where each solution entry is evaluated w.r.t. to a predefined set of dimensions.
Profit, availability and risk are examples for such abstract dimensions. Depending on
the weighting of the dimensions for a specific customer (e.g. a customer is strongly in-
terested in products with high return rates, i.e. compared to availability and risk, profit
is a very important dimension) the set of solutions is ordered using the formula g(x) =
P
n
i=1
e
i
s
i
(x), where n denotes the number of dimensions, g(x) represents the utility
of a solution x, e
i
represents the interest of the customer in dimension i, and s
i
is the
contribution of solution x to dimension i.
Presentation of Solutions. For each solution a set of immediate explanations [16] is cal-
culated, i.e. a set of explanations which are derived from variable assignments directly
dependent on selections already made during search. Furthermore, solution-specific ex-
planations are supported, e.g. if the customer is strongly interested in high return rates
and a solution shows a remarkable return rate, this fact is explicitly mentioned when the
solution is presented to the customer. In contrast to immediate explanations (derived in
the search process), solution-specific explanations are related to explicitly defined ex-
planation constraints.
171
3.3 Knowledge Acquisition
Knowledge Base Debugging. Effective debugging support for the implementation of
recommender knowledge bases is a critical issue for a successful development and
maintenance of advisors. In Koba4MS we have implemented model-based diagnosis
algorithms [13,14] supporting the identification of minimal sources of inconsistencies
in recommender knowledge bases. Similar to the diagnosis and repair of customer re-
quirements, we apply model-based diagnosis techniques in order to identify a minimal
set of constraints C which - when deleted from the knowledge base - allow consis-
tency restoration.
4 Experiences from Projects
A graphical development environment guaranteeing the maintainability of applications
is a major prerequisite for successfully implementing a knowledge-based advisor. In
the financial services domain the implementation and maintenance of knowledge bases
must be supported for non-programmers, i.e. the knowledge acquisition component
must provide intuitive modelling concepts. The overall efforts for implementing an ad-
visor application are between one man month and about a man year where the most
influencing factors are the complexity of the knowledge bases, the modeling knowledge
of people engaged in the project and efforts related to the implementation of interfaces
to remote systems (e.g. ERP or CRM systems). The correctness of solutions plays a
vital role for the acceptance of the system by sales representatives applying the system
while communicating with the customer. Domain experts do not use formal knowledge
representation formalisms on a daily basis, i.e. effective test and debugging support is
extremely useful and significantly improves the effectiveness of the overall advisor de-
velopment process. Experiences from projects indicate a reduction of efforts related to
knowledge base development of about 30-50 percent. The following conclusions can be
drawn from the actual projects based on Koba4MS technologies.
Knowledge Acquisition. Experiences from projects
4
show that graphical know-
ledge acquisition is a major precondition for enabling the design and maintenance
of recommender knowledge bases and significantly reduces the knowledge acqui-
sition bottleneck between domain experts and knowledge engineers. The usability
has been shown, e.g., in our projects in the financial services domain, where domain
experts autonomously develop and maintain the recommender knowledge bases af-
ter a first project where they were accompanied by a technical expert.
Cross Selling. Koba4MS indicates cross-selling opportunities with a corresponding
set of explanations as to why a solution is useful for the customer. The analysis of
sales records e.g. in the digital camera domain shows significant improvements in
the sales of add-on and niche products which were neglected previously.
Routine advisory tasks. Effort reductions related to routine advisory tasks are re-
ported, e.g. financial services advisory provided on the homepage relieves sales
representatives from routine advisory jobs.
4
See e.g. www.hypo-alpe-adria.at (investment advisor) or www.geizhals.at (digital camera ad-
visor deployed on the largest Austrian online product platform).
172
Documentation. Added value is provided by explanations for calculated service
portfolios which are used as starting point for future advisory sessions. Further-
more, legal regulations can force companies to provide intelligent reporting for the
customer, e.g. due to regulations of the European Union, financial service providers
are forced to improve the documentation of advisory sessions - intelligent reporting
is required which includes explanations as to why certain products were offered to
the customer.
Koba4MS knowledge bases are developed and tested by marketing and sales ex-
perts. Sales representatives can rely on the solutions calculated by the financial
advisor and can provide qualified explanations.
A set of applications has been implemented on the basis of the recommender tech-
nologies presented in this paper, e.g. the digital camera advisor PIXLA which was
implemented for the largest Austrian online product platform (www.geizhals.at).
This application exhibits about 10.000 successful advisory sessions per month.
Users of www.geizhals.at were interviewed before and after the introduction of
PIXLA. The major result of the study was a statistically significant increase of cus-
tomer satisfaction (related to dimensions such as easiness to find products etc.).
5 Conclusions
In this paper we have presented the Koba4MS toolsuite which supports the implemen-
tation of knowledge-based recommender applications (advisors). The toolsuite is based
on innovative AI technologies (model-based diagnosis, personalization, constraint sat-
isfaction) which provide an intuitive access to complex products for customers as well
as for sales representatives. Koba4MS includes a graphical development, test and de-
bugging environment which allows the development and maintenance of recommender
knowledge bases for non-programmers. The applicability of the presented concepts has
been shown within the context of commercial projects. Next steps include the integra-
tion of critiquing mechansims into the presentation of results and the application of
model-based diagnosis concepts to the identification of inconsistent transitions in rec-
ommender process definitions.
References
1. Ardissono, L., Felfernig, A., Friedrich, G., Jannach, D., Petrone, G., Schaefer, R., Zanker,
M.: A Framework for the development of personalized, distributed web-based configuration
systems. AI Magazine 24 (2003) 93–108
2. Burke, R.: Knowledge-based Recommender Systems. Encyclopedia of Library & Informa-
tion Syst. 69 (2000)
3. Burke, R.: Hybrid Recommender Systems: Survery and Experiments. User Modeling and
User-Adapted Interaction (2002) 331–370
4. Herlocker, J.L., Konstan, J.A., Terveen, L.G., Riedl, J.T.: Evaluating Collaborative Filtering
Recommender Systems. ACM Trans. on Information Systems 22 (2004) 5–53
5. Pazzani, M.: A Framework for Collaborative, Content-Based and Demographic Filtering.
Artificial Intelligence Review 13 (1999) 393–408
173
6. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., Riedl, J.: GroupLens: An Open Ar-
chitecture for Collaborative Filtering of Netnews. In: ACM 1994 Conference on Computer
Supported Cooperative Work. (1994) 175–186
7. Sarwar, B., Karypis, G., Konstan, J.A., Riedl, J.T.: Item-based collaborative filtering recom-
mendation algorithms. In: 10
th
Int. World Wide Web Conf. (2001) 285–295
8. Faltings, B., Pu, P., Torrens, M.: Designing Example-critiquing Interaction. In: International
Conference on Intelligent User Interfaces (IUI’04), Madeira, Portugal (2004) 22–29
9. McSherry, D.: Explaining the Pros and Cons of Conclusions in CBR. In: ECCBR. (2004)
317–330
10. Shih, Y.Y., Liu, D.R.: Hybrid recommendation approaches: collaborative filtering via valu-
able content information. In: 38th Hawaii International Conference on System Sciences
(HICSS’05), Big Island, Hawaii (2005) 217b
11. Felfernig, A., Friedrich, G., Jannach, D., Stumptner, M., Zanker, M.: Configuration know-
ledge representations for Semantic Web applications. AI Engineering Design, Analysis and
Manufacturing Journal 17 (2003) 31–50
12. Mittal, S., Frayman, F.: Towards a Generic Model of Configuration Tasks. In: Proceedings
11
th
International Joint Conf. on AI, Detroit, MI (1989) 1395–1401
13. Felfernig, A., Friedrich, G., Jannach, D., Stumptner, M.: Consistency-based Diagnosis of
Configuration Knowledge Bases. AI Journal 2 (2004) 213–234
14. Reiter, R.: A theory of diagnosis from first principles. Artificial Intelligence 23 (1987) 57–95
15. Felfernig, A., Friedrich, G., Jannach, D.: UML as domain specific language for the construc-
tion of knowledge-based configuration systems. IJSEKE 10 (2000) 449–469
16. Friedrich, G.: Elimination of Spurious Explanations. In: 16
th
European Conference on
Artificial Intelligence (ECAI 2004). (2004) 813–817
17. v.Noord, G., Gerdemann, D.: Finite State Transducers with Predicates and Identities. Gram-
mars 4 (2001) 263–286
18. Preece, A., Talbot, S., Vignollet, L.: Evaluation of Verification Tools for Knowledge-Based
Systems. International Journal of Human-Computer Studies 47 (1997) 629–658
19. Edvardson, J.: A Survery on Automatic Test Data Generation. In: 2
nd
Conference on Com-
puter Science and Engineering CCSSe’99. (1999)
20. Felfernig, A.: Koba4MS: Selling Complex Products and Services using Knowledge-based
Recommender Systems. In: 7th International IEEE Conference on E-Commerce Technology,
Munich, Germany (2005) to appear
21. Tsang, E.: Foundations of Constraint Satisfaction. Academic Press, London (1993)
174