Laurent Brisson
Institut TELECOM, TELECOM Bretagne, CNRS FRE 3167 LAB-STICC, Technop
ole Brest-Iroise, France
Martine Collard
INRIA Sophia Antipolis, 2004 route des Lucioles, 06902 BP93 Sophia Antipolis, France
University of Nice Sophia Antipolis, France
Data mining, Knowledge integration, Ontology Driven Information System.
This paper deals with knowledge integration in a data mining process. We suggest to model domain knowledge
during business understanding and data understanding steps in order to build an ontology driven information
system (ODIS). We present the KEOPS Methodology based on this approach. In KEOPS, the ODIS is dedi-
cated to data mining tasks. It allows using expert knowledge for efficient data selection, data preparation and
model interpretation. In this paper, we detail each of these ontology driven steps and we define a part-way in-
terestingness measure that integrates both objective and subjective criteria in order to evaluate model relevance
according to expert knowledge.
In knowledge discovery from data, methods and tech-
niques are developed for discovering specific trends
in a system or organization business by analyzing its
data. The real advantage for decision making relies on
the add-on provided by comparing extracted knowl-
edge against a priori domain knowledge. Integrating
domain a priori knowledge during the data mining
process is currently an important research issue in the
data mining field.
In this paper, we present KEOPS methodology
based on an ontology driven information system
which integrates a priori knowledge all along the data
mining process in a coherent and uniform manner. We
detail each of these ontology driven steps and we de-
fine a part-way interestingness measure that integrates
both objective and subjective criteria in order to eval-
uate model relevance according to expert knowledge.
The paper is organized in six sections. Section 2
presents the issue addressed and KEOPS main char-
acteristics. Section 3 is devoted to ontology driven in-
formation systems. In section 4, the KEOPS method-
ology is presented step by step. In Section 5, we com-
ment some results which demonstrate the relevance of
the approach. We conclude in Section 6.
The Data Mining process described according to the
CRISP-DM model (Chapman et al., 2000) is pre-
sented as both iterative and interactive. The itera-
tive nature is due to the way processes run cycling
test-error experiments. Indeed data miners have to re-
peat the pre-processing steps of domain understand-
ing, data understanding and data preparation until fi-
nal models are considered relevant. The interactive
nature is inherent to a data mining activity since com-
munications with experts is necessary for understand-
ing domain and data and for interpreting results. Is-
sues in evaluating and interpreting mining process re-
sults are currently big research challenges. In order to
avoid useless iterations on preliminary tasks and facil-
itate model interpretation, one solution is to explore
deeply expert knowledge and source data in order to
formalize them in conceptual structures and exploit
these structures both for robust data preparation and
for flexible model interpretation.
In the literature, partial solutions for domain
knowledge interpretation are proposed for optimiz-
ing pre-processing steps (Kedad and M
etais, 2002).
For model evaluation, detailed studies have been de-
Brisson L. and Collard M. (2008).
In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 54-61
DOI: 10.5220/0001697400540061
Figure 1: KEOPS methodology.
voted to interestingness measures (McGarry, 2005).
A consensus among researchers is now established
to consider objective interestingness versus subjective
interestingness. Objective interestingness is tradition-
ally evaluated by a variety of statistic indexes while
subjective interestingness is generally evaluated by
comparing discovered patterns to user knowledge or
a priori convictions of domain experts. In this pa-
per we present the KEOPS methodology based on an
ontology driven information system which addresses
the knowledge integration issue. The system relies
on three main components: an ontology, a knowl-
edge base and a mining oriented database rebuilt from
source raw data. These components allow to model
domain concepts and relationships among them. They
are used to pre-process data and to identify mapping
between discovered patterns and expert knowledge.
An ontology driven information systems is an infor-
mation system (IS) which relies mainly on an ex-
plicit ontology. This ontology may underlie all as-
pects and components of the information system. An
ODIS contains three kinds of components: applica-
tion programs, information resources and user inter-
faces. (Guarino, 1998) discusses the impact of an on-
tology on an information system according to tempo-
ral and structural dimension.
The temporal dimension refers to ontology role
during IS construction and run-time. If we have a
set of reusable ontologies, the semantic content ex-
pressed can be transformed and translated into an IS
component. Even if the volume of ontology knowl-
edge available is modest it may nevertheless help a
designer in a conceptual analysis task. This task con-
sists frequently of redesigning an existing information
system. This approach fits the needs of data mining
tasks where an operational database has to be trans-
formed into datasets before the data mining modeling
The structural dimension refers to each informa-
tion system component which may use the ontology
in a specific way.
Database component: at development time, an on-
tology can play an important role in requirement
analysis and conceptual modeling. The resulting
conceptual model can be represented as a com-
puter processable ontology mapped to a concrete
target platform (Ceri and Fraternali, 1997). Usu-
ally, IS conceptual schemes (CS) are created from
scratch, wasting a lot of time and resources.
Interface components may be assisted by ontolo-
gies which are used to generate personalized in-
terfaces or to manage user profiles (Guarino et al.,
1998; Penarrubia et al., 2004).
Application program components use implicit
knowledge in order to perform a task. However,
this knowledge is often hardcoded in software.
Ontologies may provide a formal base helping to
access domain knowledge.
KEOPS is a methodology which drives data mining
processes by integrating expert knowledge. These are
the goals addressed:
To manage interactions between knowledge and
data all along the data mining process: data prepa-
ration, datasets generation, modeling, evaluation
and results visualization.
To evaluate extracted models according to domain
expert knowledge.
To provide easy navigation throughout the space
of results.
KEOPS (cf. fig. 1) is based upon an ontology driven
information system (ODIS) set up with four compo-
An application ontology whose concepts and re-
lationships between them are dedicated to domain
and data mining task.
A Mining Oriented DataBase (MODB): a rela-
tional database whose attributes and values are
chosen among ontology concepts.
A knowledge base to express consensual knowl-
edge, obvious knowledge and user assumptions.
A set of information system components - user in-
terfaces, extraction algorithms, evaluation meth-
ods - in order to select the most relevant extracted
models according to expert knowledge.
KEOPS methodology extends the CRISP-DM
process model by integrating knowledge in most steps
of the mining process. The initial step focuses on
business understanding. The second step focuses on
data understanding and activities in order to check
data reliability. Data reliability problems are solved
during the third step of data preparation. The fourth
step is the evaluation of extracted models. In this
paper we don’t focus on modeling step of CRISP-
DM model since we ran CLOSE algorithm (Pasquier
et al., 1999) which extracts association rules without
domain knowledge.
4.1 Business Understanding
During business understanding step, documents, data,
domain knowledge and discussion between experts
lead to assess situation, to determine business objec-
tives and success criteria, and to evaluate risks and
contingencies. However this step is often rather infor-
KEOPS methodology requires to build an ontol-
ogy driven information system during the next step,
data understanding. Consequently an informal speci-
fication of business objectives and expert knowledge
is henceforth insufficient. Thus, it is necessary to for-
malize expert knowledge during business understand-
ing. We chose to state knowledge with production
rules, also called “if ... then ... rules. These rules are
modular, each defining a small and independent piece
of knowledge. Furthermore, they can be easily com-
pared to extracted association rules. Each knowledge
rule has some essential properties to select the most
interesting association rules:
Knowledge confidence level: five different values
are available to describe knowledge confidence
according to a domain expert. These values are
ranges of confidence values: 0-20%, 20-40%, 40-
60%, 60-80% and 80-100%. We call confidence
the probability for the rule consequence to occur
when the rule condition holds.
Knowledge certainty:
Obvious: knowledge cannot be contradicted.
Consensual: domain knowledge shared among
Assumption: knowledge the user wants to
Since the description of expert interview methodol-
ogy in order to capture knowledge is beyond the
scope of this paper, the reader should refer to (Becker,
4.2 Data Understanding
Data understanding means selection and description
of source data in order to capture their semantic and
reliability. During this step, the ontology is built in
order to identify domain concepts and relationships
between them (the objective is to select among data
the most interesting attributes according to the busi-
ness objectives), to solve ambiguities within data and
to choose data discretization levels.
Consequently, the ontology formalizes domain
concepts and information about data. This ontology
is an application ontology; it contains the essential
knowledge in order to drive data mining tasks. On-
tology concepts are related to domain concepts, how-
ever relationships between them model database rela-
tionships. During next step, data preparation (cf. sec-
tion 4.3), a relational database called Mining Oriented
DataBase (MODB) will be built.
In order to understand links between the MODB
and the ontology it is necessary to define notions of
domain, concept and relationships:
Domain: This notion in KEOPS methodology,
refers to the notion of domain in relational theory.
ICEIS 2008 - International Conference on Enterprise Information Systems
Figure 2: Bookshop ontology snapshoot.
A domain represents a set of values associated to
a semantic entity (or concept).
Concept: Each concept of the ontology has a
property defining its role. There exist two classes
of concepts: attribute concepts and value con-
An attribute concept is identified by a name and
a domain.
Each value of domain is called a value concept.
Thus a domain is described by an attribute
concept and by value concepts organized into a
taxonomy. Each MODB attribute is linked to
one and only one attribute concept and takes
its values in the associated domain. In figure 2
“Bookshop” is an attribute concept, Academic”
a value concept and the set {Academic, General,
Sciences, Letters} defines “Bookshop” domain.
Relationships: There exists three kinds of rela-
tionships between concepts:
A data-related relationship: “valueOf relation-
ship between an attribute concept and a value
concept. The set of value concepts linked to
an attribute concept with “valueOf relation-
ship define a domain within the MODB.
Subsumption relationship between two value
concepts. A concept subsumed by another one
is member of the same domain. This relation-
ship is useful during data preparation (to se-
lect data granularity in datasets), reduction of
rule volume (to generate generalized associa-
tion rules, see 4.4.1, comparison between mod-
els and knowledge (to consider sibling and an-
cestor concepts) and final results visualization.
Semantic relationships between value concepts.
These relationships could be order, compo-
sition, exclusion or equivalence relationships.
They can be used to compare extracted models
and knowledge and to visualize results.
KEOPS methodology aims to extract interesting mod-
els according user knowledge. Consequently, it is
necessary during ontology construction to be careful
with some usual concerns in data mining:
Aggregation level: like data, ontology concepts
have to represent disjoint domains.
Discretization level: ontology concepts have to
model various solutions for data discretization.
Bad choices may affect modeling step efficiency.
Data correlation: if concepts are strongly related
into the MODB, extracted models might be trivial
and uninteresting.
Since these concerns are beyond the scope of this
paper, the reader should refer to (De Leenheer and
de Moor, 2005) for a better insight on concept elicita-
tion and (Berka and Bruha, 1998) for a better insight
on discretization and grouping.
Table 1: Bookshop ontology concept elicitation.
Source Attribute Value
Data Concept Concept
St Denis Shop Bookshop Academic
St Denis Shop Location St Michel bd
Rive Gauche 5th Bookshop General
Rive Gauche 5th Location 5th District
Example. Let’s take the case of a bookstore com-
pany with several bookshops in Paris which plan to
improve customer relationships. Bookshops may be
specialized in a field like “academic” or not (general)
(see figure 2). Bookshops are located geographically.
Data are provided on bookshops, customers and sales.
Table 1 shows a way for mapping source values to on-
tology concepts.
4.3 Data Preparation
Data preparation is very iterative and time consuming.
The objective is to refine data: discretize, clean and
build new attributes and values in the MODB. Dur-
ing this step, KEOPS suggests building MODB by
mapping original data with ontology concepts. The
database contains only bottom ontology concepts.
The objective is to structure knowledge and data in
order to process efficient mining tasks and to save
time spent into data preparation. The idea is to al-
low generation of multiple datasets from the MODB,
using ontology relationships without another prepara-
tion step from raw data. Furthermore, during ODIS
construction, experts can express their knowledge us-
ing the ontology which is consistent with data.
4.3.1 Mining Oriented Database (MODB)
Databases often contain several tables sharing simi-
lar information. However, it is desirable that each
MODB table contains all the information semanti-
cally close and it’s important to observe normal forms
in these tables. During datasets generation, it’s easy
to use join in order to create interesting datasets to be
mined. However these datasets don’t have to observe
normal forms.
4.3.2 Datasets Generation
It’s often necessary, in a data mining process, to
step back to data preparation. Algorithms were pro-
posed to choose relevant attributes among large data
sources. However, sometimes results don’t satisfy
user expectations and datasets have to be built again
to run new tests. KEOPS methodology suggests using
the ontology in order to describe domain values and
relationships between these values. Consequently,
various datasets could be generated according to ex-
pert user choices.
The ontology driven information system allows
choosing all data preparation strategies providing var-
ious datasets from the same source values. A dataset
is built using the following operators:
Traditional relational algebra operators: projec-
tion, selection and join.
Data granularity: this operator allows choosing,
among ontology, concepts which will be in the
mining oriented database.
In order to generate datasets we developed soft-
ware whose inputs are MODB and user parameters
and outputs are new datasets. The user can graphi-
cally select relational algebra operator and data gran-
ularity. While database attributes and values are also
ontology concepts KEOPS methodology and KEOPS
software make easier the data preparation task.
4.4 Evaluation
This step assesses to what extent models meet the
business objectives and seeks to determine if there is
some business reason why these models are deficient.
Furthermore, algorithms may generate lots of mod-
els according to parameters chosen for the extraction.
That’s why evaluation is an important task in KEOPS
methodology in order to select the most interesting
models according to expert knowledge.
4.4.1 Rule Volume Reduction
We choose an association rule extraction algorithm
which can generate bases containing only minimal
non-redundant rules without information loss. Then,
these rules are filtered to suppress semantic redun-
dancies. KEOPS methodology is based on Srikant’s
generalized association rules definition (Srikant and
Agrawal, 1995). These rules are minimal because
they forbid all irrelevant relationships within their
items. We give a formal definition below:
Let T be a taxonomy of items. R : A C is called
generalized association rule if:
No item in C is an ancestor of any item in A or C
No item in A is an ancestor of any item in A
Consequently relationships appearing within
these rules are semantic and generalization relation-
ships from C items to A items. The objective is to
maximize information level in minimal rules. The
last step consists of replacing a set of these rules by a
more generalized one.
4.4.2 Rule Interestingness Evaluation
KEOPS methodology suggests comparing extracted
rules with expert’s knowledge. Extracted rules having
one or more items that are in relationship with some
knowledge rules items (i.e. value concepts are linked
in the ontology) have to be selected. Then, for each
pair knowledge rule/extracted rule:
Extracted rule antecedant coverage is compared
to knowledge rule antecedent coverage, then ex-
tracted rule consequent coverage is compared to
knowledge rule consequent coverage.
ICEIS 2008 - International Conference on Enterprise Information Systems
Table 2: Interestingness measure if confidence levels are similar.
Kind of knowledge Rule R informative level
More than K Similar Less than K
Obvious weak none none
Consensual medium weak weak
Assumption strong medium medium
By coverage comparison the most informative
rule is deduced, i.e. rule predicting the largest
consequence from the smallest condition.
IMAK interestingness measure is applied (Bris-
son, 2007). This measure evaluates rule qual-
ity considering relative confidence values, relative
information levels and knowledge certainty (see
section 4.1).
Thus, ontology driven information systems are use-
ful in order to formalize domain concepts, to ex-
press knowledge, to generate models and to facilitate
knowledge and models ontology-based comparison.
Example. Let us assume that a domain expert
makes the following assumption: “If a student wants
to buy a book about JAVA he comes to an academic
bookshop. and gives it a 60%-80% estimation of
confidence. Let us assume that the extracted rule is
slightly different because it says that “Every young
customer buying a book about J2EE comes to an aca-
demic bookshop” and has 75% confidence.
Assumption K. book=’JAVA buyer=’student’
Extracted Rule R. book=’J2EE’ buyer=’youngs’
According KEOPS methodology these two rules
are said to be comparable because at least one ex-
tracted rule item is in relationship with a knowledge
rule item: ’youngs’ is more general than student’ and
’JAVA is more general than ’J2EE’. Then, the algo-
rithm compares the coverage of these two rules in or-
der to evaluate the more informative one. Let’s make
the assumption that R is more informative than K.
Since these two rules have similar confidence we can
use table 2 in order to evaluate extracted rule interest-
ingness (similar tables for various confidence levels
are presented in (Brisson, 2007)). While the knowl-
edge is an assumption, the interestingness degree of
the extracted rule is strong.
Although we illustrated in this paper the KEOPS
methodology with bookstore example, we run exper-
iments on real data provided by French Family Al-
lowance Office (CAF: Caisses d’allocations famil-
iales). In this section we don’t express some specific
rules about allowance beneficiaries behavior (because
of privacy) but only extracted rules reliability. These
results show we are able to select relevant rules to pro-
vide to experts for final human evaluation.
CAF data were extracted during 2004 in the town
of Grenoble (France). Each row describes one contact
between the office and a beneficiary with 15 attributes
and data about 443716 contacts were provided. We
ran CLOSE algorithm and extracted 4404 association
rules. The interestingness measure, IMAK, helps to
filter the best ones. Figure 5 plots 4404 rules accord-
ing to confidence and lift.
Experiments illustrated by figure 3 and 4 com-
pare these rules to a specific knowledge. We may ob-
serve that among all of the extracted rules only few
of them are selected. Selection condition is to match
the knowledge and to have an interestingness value
greater than 0. In these figures interestingness value
is illustrated by the dot size.
In figure 3 lift of selected rules is greater than 1
and often greater than knowledge lift (lift equals 1
at independency). Furthermore, some extracted rules
have a better confidence but a smaller support: they il-
lustrated the discovery of rare events which could be
very interesting for expert users.
Figure 4 shows some results for another specific
knowledge. We may observe again that only few rules
are selected. These rules offer various tradeoff to se-
lect rare events (low support and high confidence) or
general rules (high support and good confidence) to
provide to domain experts.
As future work, we plan to evaluate rules selected
by KEOPS software with the help of some expert
groups who are able to validate their semantic rele-
Figure 3: Extracted rules (dots) matching knowledge rule 335 (square) (IMAK interestingness value increase with dot size).
a) Confidence vs Lift - b) Confidence vs Support.
Figure 4: Extracted rules (dots) matching knowledge rule 565 (square) (IMAK interestingness value increase with dot size).
a) Confidence vs Lift - b) Confidence vs Support.
Managing domain knowledge during the data min-
ing process is currently an important research issue
in the data mining field. In this paper, we presented
the so-called KEOPS methodology for integrating ex-
pert knowledge all along the data mining process in a
coherent and uniform manner.
We built an ontology driven information system
(ODIS) based on an application ontology, a knowl-
edge base and a mining oriented database rebuilt from
source raw data. Thus, expert knowledge is used dur-
ing business and data understanding, data preparation
and model evaluation steps. We show that integrat-
ing expert knowledge during the first step, gives ex-
perts a best insight upon the whole data mining pro-
cess. In the last step we introduced IMAK, a part-way
interestingness measure that integrates both objective
and subjective criteria in order to evaluate models rel-
evance according to expert knowledge.
We developed KEOPS software in order to run
experiments. Experimental results show that IMAK
measure helps to select a reduced rule set among data
mining results. These rules offer various tradeoff al-
lowing experts to select rare events or more general
rules which are relevant according to their knowledge.
ICEIS 2008 - International Conference on Enterprise Information Systems
Figure 5: Confidence vs Lift of all of the extracted rules.
Becker, H. S. (1976). Sociological Work: Method and Sub-
stance. Transaction Publishers, U. S.
Berka, P. and Bruha, I. (1998). Discretization and grouping:
Preprocessing steps for data mining. In PKDD, pages
Brisson, L. (2007). Knowledge extraction using a concep-
tual information system (excis). In Ontologies-Based
Databases and Information Systems, volume 4623 of
Lecture notes in computer science, pages 119 134,
Berlin, Heidelberg. Springer.
Ceri, S. and Fraternali, P. (1997). Designing Database Ap-
plications with Objects and Rules: The IDEA Method-
ology. Series on Database Systems and Applications.
Addison Wesley.
Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz,
T., Shearer, C., and Wirth, R. (2000). Crisp-dm 1.0:
Step-by-step data mining guide. In SPSS Inc.
De Leenheer, P. and de Moor, A. (2005). Context-driven
disambiguation in ontology elicitation. In Shvaiko, P.
and Euzenat, J., editors, Context and Ontologies: The-
ory, Practice and Applications, pages 17–24, Pitts-
burgh, Pennsylvania. AAAI, AAAI Press.
Guarino, N. (1998). Formal Ontology in Information
Systems. IOS Press, Amsterdam, The Netherlands.
Amended version of previous one in Proceedings
of the 1st International Conference June 6-8, 1998,
Trento, Italy.
Guarino, N., Masolo, C., and Vetere, G. (1998). Ontoseek:
Using large linguistic ontologies for gathering infor-
mation resources from the web. Technical report,
Kedad, Z. and M
etais, E. (2002). Ontology-based data
cleaning. In NLDB ’02: Proceedings of the 6th Inter-
national Conference on Applications of Natural Lan-
guage to Information Systems-Revised Papers, pages
137–149, London, UK. Springer-Verlag.
McGarry, K. (2005). A survey of interestingness measures
for knowledge discovery. Knowl. Eng. Rev., 20(1):39–
Pasquier, N., Bastide, Y., Taouil, R., and Lakhal, L. (1999).
Closed set based discovery of small covers for asso-
ciation rules. In Actes des 15mes journes Bases de
Donnes Avances (BDA’99), pages 361–381.
Penarrubia, A., Fernandez-Caballero, A., Gonzalez, P.,
Botella, F., Grau, A., and Martinez, O. (2004).
Ontology-based interface adaptivity in web-based
learning systems. In ICALT ’04: Proceedings of the
IEEE International Conference on Advanced Learn-
ing Technologies (ICALT’04), pages 435–439, Wash-
ington, DC, USA. IEEE Computer Society.
Srikant, R. and Agrawal, R. (1995). Mining generalized as-
sociation rules. In VLDB ’95: Proceedings of the 21th
International Conference on Very Large Data Bases,
pages 407–419, San Francisco, CA, USA. Morgan
Kaufmann Publishers Inc.