Attribute Value Ontology
Using Semantics in Data Mining
Tomasz Łukaszewski, Joanna J´ozefowska and Agnieszka Ławrynowicz
Poznan University of Technology, Poznan, Poland
Keywords:
Ontologies, Imperfect Data, Na¨ıve Bayesian Classifier.
Abstract:
We propose a new concept to represent attribute values as an ontology that allows modeling different levels
of abstraction. In this way more or less precise values may be used instead of missing or erroneous data. The
goal is to use this representation in order to improve analysis of imperfect data. The proposed attribute value
ontology (AVO) allows to upgrade the precision of information not only from positive observations but also
from negative ones. We show how to classify a new example using a set of training examples described in the
same or more precise way. Another advantage of the proposed approach is providing an efficient way to avoid
the effect of overfitting.
1 INTRODUCTION
The concept of using ontologies in order to enhance
data with semantics resulted in a new vision of com-
puting - semantic computing. On the one hand, se-
mantic computing allows to integrate heterogeneous
data sources. On the other hand, semantic computing
allows to improve the data analysis e.g. the analysis
of imperfect data.
Imperfect data (e.g. erroneous or missing attribute
values) are very common in the field of Data Min-
ing and they have a negative effect on the mining re-
sults. Let us notice, that some erroneous or missing
attribute values may be introduced by users that are
required to provide very specific values, but the level
of their knowledge of the domain is only very general
and they are unable to precisely describe the obser-
vation by an appropriate value of an attribute. Even
if a person is an expert in the domain, erroneous or
missing attribute values can be introduced as a conse-
quence of lack of time or other resources to precisely
describe the observation by an appropriate value.
In this paper we present a semantic approach to
Data Mining (Witten et al., 2011) aiming at improve-
ment of the analysis of imperfect data using some
background knowledge. Most common approach to
exploit background knowledge in data mining is gen-
eralization of attribute values (Han et al., 1992). It
allows to define abstract concepts as generalizations
of the primitive ones. Background knowledge used in
this approach has a form of taxonomies, categories or
more general relationships between attributes.
We introduce an attribute value ontology (AVO)
in order to improve the expressiveness of the knowl-
edge representation language. Firstly, AVO allows
to model different levels of abstraction - precise and
imprecise values. As a result, users are allowed to
use (less or more) imprecise values instead of erro-
neous or missing values. Secondly, AVO allows to
precise values not only by indicating the subconcept
of a current concept (a positive observation) but also
rejecting some subconcepts of the current concept (a
negative observation). This is a common technique
of describing what something is by explaining what
it is not. Moreover, AVO allows to avoid overfitting
by analysing results not only for a current value but
also for less precise values. This approach is an anal-
ogy to the analysis of results for special cases (ex-
ceptions) and for general ones. Finally, we present a
na¨ıve Bayesian classifier extended to AVO.
2 ATTRIBUTE VALUE
ONTOLOGY
Let us assume that given is an ontology, which rep-
resents the domain knowledge. In particular, it ex-
presses a multilevel subsumption hierarchy of con-
cepts (ISA hierarchy) representing different levels of
abstraction - precise and imprecise values. We define
an attribute value ontology (AVO) as follows:
329
Łukaszewski T., Józefowska J. and Ławrynowicz A..
Attribute Value Ontology - Using Semantics in Data Mining.
DOI: 10.5220/0004157003290334
In Proceedings of the 14th International Conference on Enterprise Information Systems (SCOE-2012), pages 329-334
ISBN: 978-989-8565-11-2
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
Definition 1. An attribute value ontology (AVO) is
a pair A = hC, Ri, where: C is a set of concepts (prim-
itive and abstract ones), R is a subsumption relation
over C, subset C
P
⊆ C of concepts without predeces-
sors is a finite set of primitive concepts of A.
For simplicity of presentation we consider the hi-
erarchy of concepts such that each concept has at most
one predecessor (tree structure). In general concepts
may have multiple predecessors (DAG structure).
2.1 Levels of Abstraction
Given is an attribute A and a set V = {v
1
, v
2
, ..., v
n
},
n > 1, of specific values of this attribute. These spe-
cific values can be interpreted as a single level of ab-
straction. The user that is unable to precisely describe
the observation by a specific value of A has to manage
with a missing or erroneous value.
Introducing AVO we add new levels of abstrac-
tion. We assume that primitive concepts of AVO rep-
resent these specific values of A (the original level
of abstraction). Abstract concepts of AVO are used
when the users are unable to precisely describe the
observation by a specific value of A (the new levels
of abstraction). Therefore, we call primitive and ab-
stract concepts: precise and imprecise values of AVO,
respectively.
Infectious Agent
Bacteria
Gram-positive
Bacteria
Streptococcous
Gram-negative
Bacteria
E.Coli Salmonella
Fungi Virus
Figure 1: Example of an attribute value ontology.
Example 1. Let us consider the following medical
problem. In order to determine the correct treatment,
an agent that caused the infection needs to be speci-
fied. Although, all viral infections determine the same
treatment (similarly infections cased by fungi), iden-
tification of the bacteria type is important in order to
decide about the appropriate treatment. Thus, specific
values of this attribute are the following: Streptococ-
cus, E.Coli, Salmonella, Fungi, Virus. An AVO de-
scribing the domain of infectious agents is presented
in Fig. 1. Primitive concepts of AVO represent these
specific values. Abstract concepts are the following:
Infectious Agent, Bacteria, Gram-positive Bacteria,
Gram-negative Bacteria.
A user that is unable to precisely describe the in-
fectious agent can use the most abstract concept of
this AVO (Infectious Agent) or one of its abstract sub-
concepts (Bacteria, Gram-positive Bacteria, Gram-
negative Bacteria). In the next subsection we show
that AVO allows to precise values not only by indi-
cating the subconcept of a current concept but also
rejecting some subconcepts of the current concept.
Let us observe that Streptococcus is not the only
Gram-positive Bacteria in the real world and our hi-
erarchy, for some reasons, does not contain concepts
of the other Gram-positive Bacteria. In order to
represent this we make the open world assumption
(OWA). Therefore, the concept Gram-positive Bacte-
ria should be correctly interpreted as: Streptococcus
or other Gram-positive Bacteria. Similarly, the con-
cept Gram-negative Bacteria should be interpreted as:
E.Coli or Salmonella or other Gram-negative Bacte-
ria. In the next subsection we show the consequences
of assuming the open world.
2.2 Increasing the Precision
In the previous section we have shown that the user
that is unable to precisely describe the observation
may use abstract concepts (imprecise values) of AVO.
A person that knowsnothing can use the most abstract
concept of AVO. Considering our medical problem,
the user shall use the abstract concept Infectious Agent
- Fig. 1.
AVO allows to precise descriptions explicitly, in-
dicating the subconcept of a current concept. Con-
sidering our medical problem, let us assume, that the
user has made a positive observation that this infec-
tious agent is Bacteria - Fig. 2.
Infectious Agent
Bacteria
Gram-positive
Bacteria
Streptococcous
Gram-negative
Bacteria
E.Coli Salmonella
Fungi Virus
Figure 2: Positive observation: infectious agent is bacteria.
AVO allowsto precise descriptions also implicitly,
rejecting some subconcepts of a current concept. This
is a common technique of describing what something
is by explaining what it is not. Continuing or medi-
cal example, let us assume, that the user has made a
negative observation that this Bacteria is not Gram-
negative Bacteria - Fig. 3.
Let us notice, that making this negative observa-
tion, all but one subconcepts of Bacteria have been
ICEIS2012-14thInternationalConferenceonEnterpriseInformationSystems
330
rejected. Making the closed world assumption (CWA)
we would be allowed to say automatically, that Bac-
teria is Gram-positive Bacteria. However, we have
made the open world assumption (OWA) and we are
not allowed to precise the description in this way.
Infectious Agent
Bacteria
Gram-positive
Bacteria
Streptococcous
Gram-negative
Bacteria
E.Coli Salmonella
Fungi Virus
Figure 3: Negative observation: bacteria is not gram-
negative bacteria.
2.3 Avoidance of Overfitting
Many Data Mining algorithms construct rules or trees
that are overfitted to the training data. The best
way to avoid overfitting is to simplify the learned
model (Witten et al., 2011). This simplification can
be reached by generating sensible rules or pruning
trees. Both approaches are aimed at rejecting rules
that cover too few training examples.
AVO allows to avoid overfitting by analysing re-
sults not only for a current level of precision but also
for less precise values (superconcepts of the current
concept). This approach is an analogy to the analysis
of results for a special case (an exception) and for the
general case.
3 EXTENDING NA
¨
IVE BAYESIAN
CLASSIFIER BY AVO
Using AVO we are able to represent training and
testing examples with precise and imprecise descrip-
tions. In this section we show how to extend the na¨ıve
Bayesian classifier by AVO in order to learn this clas-
sifier from precisely and/or imprecisely described ex-
amples and classify precisely and/or imprecisely de-
scribed examples.
3.1 Na
¨
ıve Bayesian Classifier
The most straightforward and widely tested method
for probabilistic induction is known as the na¨ıve
Bayesian classifier. Despite its simplicity and
the strong conditional independence assumptions it
makes, the na¨ıve Bayesian classifier often performs
remarkably well, competitively with other well-
known induction techniques such as decision trees
and neural networks. The na¨ıve Bayesian classifier
is often used for classification problems, in which a
learner attempts to construct a classifier from a given
set T of training examples with class labels.
Assume that given is a set of n attributes
A
1
, A
2
, ..., A
n
. An (training or testing) example is rep-
resented by a vector (v
1
, v
2
, ..., v
n
), where v
i
is the spe-
cific value of A
i
. LetC represent the class variable and
C
j
represent the value it takes (a class label).
The Bayesian classifier (and also the na¨ıve
Bayesian classifier) is a classification method, which
classifies a new observation E by selecting the class
C
j
with the largest posterior probability P(C
j
|E), as
indicated below:
P(C
j
|E) =
P(C
j
)P(E|C
j
)
P(E)
. (1)
P(E) is ignored, since it is the same for all classes,
and does not affect the relative values of their proba-
bilities:
P(C
j
|E) ∝ P(C
j
)P(E|C
j
) . (2)
Since E is a composition of n discrete values, one
can expand this expression:
P(C
j
|v
1
, v
2
, ..., v
n
) ∝ P(C
j
)P(v
1
, v
2
, ..., v
n
|C
j
) . (3)
where P(v
1
, v
2
, ..., v
n
|C
j
) is the conditional prob-
ability of the example E given the class C
j
; P(C
j
) is
the prior probability that one will observe class C
j
.
All these parameters are estimated from the training
set. However, a direct application of these rules is
difficult due to the lack of sufficient data in the train-
ing set to reliably obtain all the probabilities needed
by the model. The na¨ıve Bayesian classifier assumes
that the attributes are conditionally independent given
the class variable, which gives us:
P(C
j
|v
1
, v
2
, ..., v
n
) ∝ P(C
j
)
∏
i
P(v
i
|C
j
) . (4)
P(v
i
|C
j
) is the probability of an instance of class
C
j
having the observedattributeA
i
value v
i
. The prob-
abilities in the above formula must be estimated from
training examples, e.g. using relative frequency:
P(C
j
) =
n
j
n
P(v
i
|C
j
) =
n
ij
n
j
(5)
where n is the number of training examples, n
j
is the
number of training examples with class label C
j
, n
ij
is the number of training examples with the value of
the attribute A
i
= v
i
and class label C
j
.
AttributeValueOntology-UsingSemanticsinDataMining
331
3.2 Inference with AVO
In the proposed approach with abstract concepts (with
AVO), the na¨ıve Bayesian classifier needs to be gen-
eralized to estimate P(c
i
|C
j
), where c
i
is a primitive
or an abstract concept of A
i
. Let us recall, that for a
given concept c
i
in A
i
, all the concepts that are more
specific than the concept c
i
are the descendants of this
concept c
i
. In order to estimate P(c
i
|C
j
), e.g. by rela-
tive frequency, we use the following property:
P(c
i
|C
j
) =
n
ij
+
∑
c
k
∈desc(c
i
)
n
kj
n
j
(6)
where n
j
is the number of training examples with
class label C
j
, n
ij
is the number of training examples
with the value of the attribute A
i
= c
i
and class la-
bel C
j
, n
kj
is the number of training examples with
the value of the attribute A
i
= c
k
and class label C
j
,
desc(c
i
) is the set of concepts that are descendants of
the concept c
i
.
The proposed approach is a generalization of the
classical approach (without abstract concepts). In the
classical approach, specific attribute values can be in-
terpreted as a single level of knowledge granularity,
and a new example is classified using training ex-
amples described by the same specific attribute value
only. In the proposed approach (with abstract con-
cepts) each descendant of a given concept ’is’ this
concept. Therefore, in the classification of a new ex-
ample described by a concept c
i
we use also all train-
ing examples described by descendants of c
i
.
The algorithm has been implemented and tested
on a series of examples available in the literature. The
results are promising, although more extensive exper-
iments are necessary in order to present statistically
valid conclusions.
3.3 Illustrative Example
Let us consider the medical problem presented in Ex-
ample 1. In order to determine the correct treatment,
an agent that caused the infection needs to be spec-
ified. For the simplicity of the presentation we con-
sider only one attribute. The training data is given
in Table 1. The first column of the table presents
the number of training examples described by a given
value (precise or imprecise) of the infectious agent for
each class. For example, the imprecise value Bacteria
is used in order to describe 6 instances with the class
label C
1
and 7 instances with the class label C
2
. In the
considered example each class is described exactly
by the same number of instances, therefore the prior
probability that one will observe class C
j
is equal to
0.5 for C
1
and C
2
.
Table 1: A medical diagnosis training data.
Instances Infectious Agent Class
6 Bacteria C1
3 Gram-positive Bacteria C1
1 Gram-negative Bacteria C1
7 Bacteria C2
1 Streptococcus C2
2 Gram-negative Bacteria C2
Infectious Agent is not Known. Let us consider
the following scenario: there is a patient and the di-
agnosis is not known. Knowing nothing we are al-
lowed to use the most abstract concept of AVO - In-
fectious Agent. We estimate the posterior probabil-
ity P(C
j
|InfectiousAgent). Therefore, we concen-
trate on these examples, that are described by the
value Infectious Agent or its descendants. From (6)
we have: P(In fectiousAgent|C
1
) =
0+(6+3+1)
10
= 1,
P(Inf ectiousAgent|C
2
) =
0+(7+1+2)
10
= 1. From (4)
we have: P(C
1
|Inf ectiousAgent) ∝ 0.5∗ 1 = 0.5, and
P(C
2
|Inf ectiousAgent) ∝ 0.5 ∗ 1 = 0.5. As we can
see, both classes are equally probable.
Infectious Agent is Bacteria. Let us assume, that
the user has made a positive observation that this
infectious agent is Bacteria - Fig. 2. We esti-
mate the posterior probability P(C
j
|Bacteria). There-
fore, we concentrate on these examples, that are
described by the value Bacteria or its descendants.
From (6) we have: P(Bacteria|C
1
) =
6+(3+1)
10
= 1,
P(Bacteria|C
2
) =
7+(1+2)
10
= 1. From (4) we have:
P(C
1
|Bacteria) ∝ 0.5∗1= 0.5, and P(C
2
|Bacteria) ∝
0.5 ∗ 1 = 0.5. As we can see, both classes are still
equally probable.
Bacteria is not Gram-negative Bacteria. Let us as-
sume, that the user has made a negative observation
that Bacteria is not Gram-negative Bacteria (shortly:
Bacteria¬GnB) - Fig. 3. We estimate the pos-
terior probability P(C
j
|Bacteria¬GnB). There-
fore we again concentrate on these examples, that are
described by the value Bacteria or its descendants.
However, this time we do not take into the consid-
eration these training examples, that are described by
the concept Gram-negative Bacteria and its subcon-
cepts. From (6) we have: P(Bacteria¬GnB|C
1
) =
6+(3)
10
= 0.9, P(Bacteria¬GnB|C
2
) =
7+(1)
10
= 0.8.
From (4) we have: P(C
1
|Bacteria¬GnB) ∝ 0.5 ∗
0.9 = 0.45, and P(C
2
|Bacteria¬GnB) ∝ 0.5 ∗
0.8 = 0.4. As we can see, there is a small difference
between the probabilities of these classes. Therefore,
additional observations are recommended.
Avoidance of Overfitting. Let us assume, that
the user has made a positive observation that Bac-
ICEIS2012-14thInternationalConferenceonEnterpriseInformationSystems
332
teria is Streptococcus. We estimate the poste-
rior probability P(C
j
|Streptococcus). Therefore
we concentrate on these examples, that are de-
scribed by the value Streptococcus or its descendants.
From (6) we have: P(Streptococcus|C
1
) =
0+(0)
10
=
0, P(Streptococcus|C
2
) =
1+(0)
10
= 0.1. From (4)
we have: P(C
1
|Streptococcus) ∝ 0.5 ∗ 0 = 0, and
P(C
2
|Streptococcus) ∝ 0.5∗ 0.1 = 0.05.
The result indicates that only class C2 is probable.
However, this result is based on a very small set of
training examples. Therefore, this result should be
treated rather as the exception from a general rule.
In our example the general rule indicates that both
classes are equally probable. However, awareness of
this exception may be very valuable in some cases.
4 RELATED WORK
Data Mining with background knowledge has been
extensively studied in the past. One of the aspects
of the background knowledge are relations between
the attribute values. Generalization of attribute val-
ues is the simplest relation considered in this context.
It allows to get abstract concepts as generalizations
of the primitive ones. Background knowledge used
in this approach has a form of taxonomies, categories
or more general relationships between concepts. Ab-
stract concepts are used in the data mining tasks in
various ways.
Compactness and Generality of Results. In the
early approaches generalization was carried out in or-
der to get more compact and more general data min-
ing results. Two groups of methods may be distin-
guished in this line. The first group consists of meth-
ods where abstract concepts replace the data values
in the original database before applying the core data
mining algorithm. This approach is used, for exam-
ple, in: (Walker, 1980), (Han et al., 1992) and (Ku-
doh et al., 2003). In the methods of the second group
generalization is integrated with the data mining al-
gorithm. Among others, this approach was applied
in: (N´u˜nez, 1991), (Almuallim et al., 1996), (Tanaka,
1996), (Taylor et al., 1997). In (N´u˜nez, 1991) an al-
gorithm EG2 (Economic Generalizer 2) was proposed
to build a decision tree. The background knowledge
contains ISA hierarchies of attribute values. At each
node of the decision tree, this algorithm builds a union
of abstract values and primitive values. In (Almual-
lim et al., 1996) an algorithm was proposed to find a
multiple-split test on hierarchical attributes (ISA hi-
erarchies) in decision tree learning. The proposed
multiple-split test is a cut through a hierarchy, which
maximizes the gain-ratio measure. The idea of cut
was proposed in (Haussler, 1988). The cut through
a hierarchy allows to reduce multiple levels of ab-
straction to a single level of abstraction and apply
classical algorithms. However, the number of possi-
ble cuts (split tests) grows exponentially in the func-
tion of leaves of the hierarchy. It turns out that this
task is very similar to the task of decision tree prun-
ing and this allows to employ a decision tree prun-
ing technique introduced in (Breiman et al., 1984). In
(Tanaka, 1996) a very similar approach was proposed
to build decision trees using structured attributes (ISA
hierarchies) and called LASA (LearniAVT-DTng Al-
gorithm with Structured Attributes). This approach
defines the unique and complete cover node set which
corresponds to the cut through a hierarchy. A mea-
sure of generalization goodness was proposed, which
takes into account two mutually conflicting factors: a
generalization level and a penalty for the induced er-
rors. An algorithm to find optimum generalization,
that transforms the original problem to the shortest
path problem, was also proposed. A simple experi-
ment showed, that the classification results of the pro-
posed approach are better than a standard approach in
terms of classification accuracy. (Taylor et al., 1997)
applied a tool ParkaDB to integrate databases and on-
tologies in order to generate classification rules based
on generalized concepts from an ontology. The level
of the generalization is determined by gathering fre-
quency counts and evaluating so called strong indica-
tors for class membership.
Handling Imprecise Descriptions. A more recent
approach is to use abstract concepts (imprecise val-
ues) in order to represent real objects that can not
be precisely described by the available specific val-
ues of an attribute. The use of attribute value tax-
onomies (AVT) in the decision tree learning (AVT-
DTL) and the na¨ıve Bayesian classifier (AVT-NBL)
is presented respectively in (Zhang et al., 2002) and
(Zhang et al., 2006). AVT-DTL and AVT-NBL, to
the best of our knowledge, are the only one existing
approaches for learning classifiers from imprecisely
specified instances and classifying imprecisely spec-
ified instances. AVT-DTL and AVT-NBL use a cut
through a hierarchy of concepts. When instances are
described by abstract values below the cut through a
taxonomy, they are aggregated upwards and stored in
abstract values of the cut. When they are described by
abstract values above the cut, they are propagated to
their descendants in the cut, proportionally.
Our approach (na¨ıve Bayesian classifier extended to
AVO) is an improvement of the AVT-DTL and AVT-
NBL in three directions. Firstly, AVT-DTL and AVT-
NBL use a cut through a taxonomy. This cut reduces
the considered taxonomy to a single level of abstrac-
AttributeValueOntology-UsingSemanticsinDataMining
333
tion only. This reduction was necessary in order to
apply a classical algorithm of decision tree learning
that was designed for a single level of abstraction.
The use of this cut in AVT-NBL is just a tradeoff be-
tween the complexity and accuracy of the classifier
(Zhang et al., 2006). We show that na¨ıve Bayesian
classifier can be extended to AVO without any reduc-
tion of a given ontology. Secondly, the semantics of
AVO allows to precise descriptions explicitly (posi-
tive observations) and implicitly (negative observa-
tions). Thirdly, the semantics of AVO allows to avoid
overfitting in a very effective way by analyzing results
for less precise values. The overfitting avoidance is a
very important issue in Data Mining and is very im-
portant from the practical point of view.
5 CONCLUSIONS
In this paper we proposed an extension of the na¨ıve
Bayesian classifier by an attribute value ontology
(AVO) aiming at the improvement of the analysis of
imperfect data. In the proposed approach, every at-
tribute is a hierarchy of concepts from the domain
knowledge base (ISA hierarchy). This semantic ap-
proach to Data Mining allows to describe examples
either very precisely or, when it is not possible, in a
more general way (using a concept from higher levels
of the hierarchy). As a result, users that are unable to
precisely describe the observation by a specific value
of an attribute, are allowed to use (less or more) im-
precise values.
Let us notice, that each imprecise value of AVO,
except the most abstract concept, is more precise than
the missing value, represented by this most abstract
concept. Therefore, introducing these abstract con-
cepts we improve the analysis of imperfect data. This
improvement is increased by each upgrade of the pre-
cision of information. We showed that even negative
observations improve this precision: ”knowing what
we do not know” is already information.
We could ask a question: how far should we pre-
cise the description? There is no single answer for
this question. Each cut through a hierarchy seems to
be a tradeoff between the complexity and accuracy.
Therefore, maintaning all the levels of abstraction is
an alternative approach to this problem. It allows to
compare results for many levels of abstraction. That
can be an efficient way to avoid the effect of overfit-
ting. Moreover, this comparison can be utilized by
a cost sensitive computing. High precision carries a
high cost. The challenge is to exploit the tolerance
for imprecision. Further research aims at experimen-
tal evaluation of the proposed approach.
ACKNOWLEDGEMENTS
Research supported by the Polish Ministry of Science
and Higher Education grant No. N N516 186437.
REFERENCES
Almuallim, H., Akiba, Y., and Kaneda, S. (1996). An
efficient algorithm for finding optimal gain-ratio
multiple-split tests on hierarchical attributes in deci-
sion tree learning. In Proceedings of the Thirteenth
National Conference on Artificial Intelligence. AAAI
Press.
Breiman, L., Friedman, J. H., Olshen, R. A., and Stone,
C. J. (1984). Classification and Regression Trees.
Wadsworth, Belmont, California, 3rd edition.
Han, J., Cai, Y., and Cercone, N. (1992). Knowledge dis-
covery in databases: An attribute-oriented approach.
In Proceedings of the 18th International Conference
on Very Large Data Bases. Morgan Kaufmann.
Haussler, D. (1988). Quantifying inductive bias: Ai learn-
ing algorithms and valiant’s learning framework. In
Artif. Intell., Vol. 36(2). Elsevier.
Kudoh, Y., Haraguchi, M., and Okubo, Y. (2003). Data
abstractions for decision tree induction. In Theoretical
Computer Science, Vol. 292(1). Elsevier.
N´u˜nez, M. (1991). The use of background knowledge in
decision tree induction. In Machine Learning, Vol.
6(3). Springer.
Tanaka, H. (1996). Decision tree learning algorithm with
structured attributes: Application to verbal case frame
acquisition. In Proceedings of the 16th International
Conference on Computational Linguistics. Center for
Sprogteknologi, Danmark.
Taylor, M. G., Stoffel, K., and Hendler, J. A. (1997).
Ontology-based induction of high level classification
rules. In Proceedings of the Workshop on Research
Issues on Data Mining and Knowledge Discovery.
ACM.
Walker, A. (1980). On retrieval from a small version of a
large data base. In Proceedings of the Sixth Interna-
tional Conference on Very Large Data Bases. IEEE
Computer Society.
Witten, I., Frank, E., and Hall, M. (2011). Data Mining.
Practical Machine Learning Tools and Techniques.
Morgan Kaufmann, Burlington, 3rd edition.
Zhang, J., Kang, D., Silvescu, A., and Honavar, V. (2006).
Learning accurate and concise naive bayes classifiers
from attribute value taxonomies and data. In Knowl.
Inf. Syst., Vol. 9(2). Springer.
Zhang, J., Silvescu, A., and Honavar, V. (2002). Ontology-
driven induction of decision trees at multiple levels of
abstraction. In LNCS Vol. 2371. Springer.
ICEIS2012-14thInternationalConferenceonEnterpriseInformationSystems
334
