Journal of Information, Control and Management Systems, Vol. 4, (2006), No.

143

LEARNING FUZZY RULES FROM FUZZY DECISION

TREES
Karol MATIAKO, Jn BOHIK, Vitaly LEVASHENKO, tefan KOVALK

University of ilina, Faculty of Management Science and Informatics,

Slovak Republic
e-mail: [email protected], [email protected],
[email protected], [email protected]
Abstract
Classification rules are an important tool for discovering knowledge from databases.
Integrating fuzzy logic algorithms into databases allows us to reduce uncertainty which
is connected with data in databases and to increase discovered knowledges accuracy.
In this paper, we analyze some possible variants of making classification rules from a
given fuzzy decision based on cumulative information. We compare their classification
accuracy with the accuracy which is reached by statistical methods and other fuzzy
classification rules.

Keywords: fuzzy rules, fuzzy decision trees, classification, fuzzy logic, decision
making, machine learning
1

INTRODUCTION
Now, data repositories are growing quickly and contain huge amount of data from
commercial, scientific, and other domain areas. According to some estimates, the
amount of data in the world is doubling every twenty months [7]. Because of it,
information and database systems are widely used. The current generation of database
systems is based mainly on a small number of primitives of Structured Query
Language (SQL). The database systems usually use relational databases which consist
of relations (tables) [11]. The tables have attributes in columns and instances in rows.
Because of increasing data, these tables contain more and more covert information.
The covert information can't be found out and transformed into knowledge by classical
SQL queries any more. One approach how to extract knowledge from the tables is to
search dependences from the data.
There are a lot of ways how to describe the dependences among data: statistical
methods Nave Bayes Classifier, k-Nearest Neighbor Classifier etc., neural networks,
decision tables, decision trees, classification rules.

144

Learning Fuzzy Rules from Fuzzy Decision Trees

Nave Bayes Classifier (NBC) represents each instance as a N-dimensional vector

of attribute values [a1, a2, , aM]. Given that there are mc classes c1, , ck, , cmc,
the classifier predicts a new instance (unknown vector X = [a1, a2, , aM]) as
belonging to the class ck having the highest posterior probability conditioned on X. In
other words, X is assigned to class ck if and only if P(ck/X) > P(ck/X) for 1 j mc and
j k. NBC is very simple, it requires only a single scan of the data, thereby providing
high accuracy and speed for large databases. Their performance is comparable to that
of decision trees and neural networks. However, there arise inaccuracies due to a) the
simplified assumptions involved and b) a lack of available probability data or
knowledge about the underlying probability distribution [13, 20].
K-Nearest Neighbor Classifier (k-NN) assumes all instances correspond to points
in the n-dimensional space Rn. During learning, all points with known classes are
remembered. When a new point is classified, the k-nearest points to the new point are
found and are used with a weight for determining the new points class. For the sake of
increasing classification accuracy, greater weights are given to closer points during
classification [12]. For large k, the algorithm is computationally more expensive than a
decision tree or neural network and requires efficient indexing. When there are lots of
instances with known classes, a reduction of their number is needed.
A neural network consists of a system of interconnected neurons. A neuron, in a
simplified way, takes positive and negative stimuli (numerical values) from other
neurons and when the weighted sum of the stimuli is greater than given threshold
value, it activates itself. The neurons output value is usually a non-linear
transformation of the sum of stimuli. In more difficult models, the non-linear
transformation is adapted by some continuous functions. Neural networks are good for
classification. They are an alternative to decision trees or rules when understandable
knowledge is not required. A disadvantage is that it is sometimes difficult to determine
the optimal number of neurons [14].
A decision table is the simplest, most rudimentary way of representing
knowledge. It has instances in rows and attributes in columns. The attributes describe
some properties of the instances. One (or more if its needed) of the attributes are
considered as class attribute. Its usually the last columns attribute. The class attribute
classifies the instances into a class - an admissible value of the class attribute. A
decision table is simply makeable and useable, e. g. a matrix. On the other hand, such a
table may become extremely large for complex situations and its usually hard to
decide which attributes to leave out without affecting the final decision [15].
A group of classification rules is also a popular way of knowledge
representation. Each rule in the group usually has the following form: IF Condition
THEN Conclusion. Both Condition and Conclusion contain one or more
attribute is attributes value and there are operators among them, such as OR, AND.
All classification rules in a group have in their conclusions only the same class
attributes. A group of rules is used for classifying instances into classes [2].

Journal of Information, Control and Management Systems, Vol. 4, (2006), No. 2

145

Another knowledge representation is a (crisp) decision tree. Such a decision tree

consists of several nodes. There are two types of nodes: a) the root and internal nodes,
b) leaves. The root and internal nodes are associated with a non-class attribute.
Leaves are associated with a class. Basically, each non-leaf node has an outgoing
branch for each its attributes possible. To classify a new instance using a decision tree,
beginning with the root, successive internal nodes are visited until a leaf is reached. At
the root or each internal node, the test for the node is applied to the instance. The
outcome of the test at an internal node determines the branch traversed, and the next
node visited. The class for the instance is simply the class of the final leaf node [5].
Thus, the conjunction of all the conditions for the branches from the root to a leaf
constitutes one of the conditions for the class associated with the leaf. It implicates a
decision tree can be represented by classification rules which use AND operator.
The paper is organized as follows. Section 2 explains why its needed to integrate
fuzzy logic into FDTs and fuzzy rules. It also offers a survey of FDTs inductions
methods. Section 3 describes how to make fuzzy classification rules from FDTs based
on cumulative information and how to use them for classification. Section 4
demonstrates our studys experimental results on several databases. Section 6
concludes this work.
2

DATA IMPERFECTION AND FUZZY DECISION TREES

To be able to work with real data its needed to consider cognitive uncertainties.
Because of it, data arent often accurate and contain some uncertainty. These
uncertainties are caused by various reasons, such as a) there is no reliable way how to
measure something e. g. a companys market value, a persons value, etc., b) it can
be too expensive to measure something exactly, c) vagueness it is associated with
humans difficulty of making sharp boundaries e. g. its strange for people to say its
hot when its 30 C and its not hot when its 29.9C, d) ambiguity it is associated
with one-to-many relations, i.e., situations with two or more alternatives such that the
choice between them is left unspecified. These problems with real data have been
successfully solving thanks to fuzzy sets and fuzzy logic for several years. Fuzzy sets
are a generalization of (crisp) sets. In the case of a set, an element x either is the sets
member or isnt. In the case of a fuzzy set, an element x is the sets member with given
membership degree which is denoted by membership function (x). (x) has a value in
the continuous interval 0 to 1 [23]. Fuzzy sets are one of main ideas of Fuzzy Logic,
which is considered as an extension of Boolean Logic as well as Multi-Valued Logic
[21].
Decision trees, which make use of fuzzy sets and fuzzy logic for solving the
introduced uncertainties, are called Fuzzy decision trees (FDTs). It has been developed
several algorithms for making FDTs so far. The algorithms usually make use of
principles of ID3 algorithm which was initially proposed by Quinlan [16] and which
was meant for making crisp decision trees. The algorithms a top-down greedy
heuristic, which means that makes locally optimal decisions at each node. The

146

Learning Fuzzy Rules from Fuzzy Decision Trees

algorithm uses minimal entropy as a criterion at each node. So-called Fuzzy ID3 is a
direct generalization of ID3 to fuzzy case. Such an algorithm uses minimal fuzzy
entropy as a criterion, e. g. [19, 17]. Notice that there are several definitions of fuzzy
entropy [1]. A variant of Fuzzy ID3 is [22] which uses minimal classification
ambiguity instead of minimal fuzzy entropy. An interesting investigation to ID3-based
algorithms, which uses cumulative information estimations, was proposed in [10].
These estimations allow us to build FDTs with different properties (unordered,
ordered, stable, etc.). There are also some algorithms, which are quite different from
ID3. For example, Wang et al. [18] present optimization principles of fuzzy decision
trees based on minimizing the total number and average depth of leaves. The algorithm
minimizes fuzzy entropy and classification ambiguity at node level, and uses fuzzy
clustering so as to merge branches. Another non-Fuzzy ID3 algorithm is discussed in
[4]. The algorithm uses look-ahead method. Its goal is to evaluate so-called
classifiability of instances that are split along branches of a give node.
One way how to make fuzzy rules is a transformation of fuzzy decision trees. In a
crisp case, each leaf of the tree corresponds to one rule. The condition of the rule is a
group of attribute is attributes value which are connected with operator AND. These
attributes in the condition are attributes associated with the nodes in the path from the
root to a considered leaf. The attributes values are the values associated with outgoing
branch of a particular attributes node in the path. The conclusion of the rule is class
attribute is class. The conclusion is the output for the condition. In other words, its
classification for a known instance whose values correspond to the condition. This
notion is based on rules made from crisp decision tree [16]. In this paper in section 3,
we describe a generalization for a fuzzy case. For a fuzzy case, the class values for a
new instance (or, more exactly, membership functions values for particular classes)
are described by several rules which belong to one or more leaves.
3

FUZZY CLASSIFICATION RULES INDUCTION.

The main goal of a made group of classification rules is to determine the value of
the class attribute from the attributes values of a new instance. In this section, we
explain the mechanism of making fuzzy rules from FDTs and the mechanism of their
use for classification. To explain the mechanisms, a FDT which is made on basis of
cumulative information is considered [10].
Let the FDT be made from a database whose instances are described by N
attributes A = {A1, ..., Ai, , AN}. Let an attribute Ai take mi possible values ai,1, ai,2,
, ai,mi, which is denoted by Ai = {ai,1, ..., ai,j, , ai,mi}. In the FDT, each non-leaf
node is associated with an attribute from A. When Ai is associated with a non-leaf
node, the node has mi outgoing branches. The j-th branch of the node is associated with
value ai,j Ai. In general, the database can contain several class attributes which
depend on non-class attributes. Because of simplicity, only one class attribute C is
considered here. The class attribute C has mc possible values c1, ..., ck, , cmc, which is

Journal of Information, Control and Management Systems, Vol. 4, (2006), No. 2

147

denoted by C = {c1, ..., ck, , cmc}. Let the FDT have R leaves L = {l1, ..., lr, ..., lR}. It
is also considered there is a value Fkr for each r-th leaf lr and each k-th class ck. This
value Fkr means the certainty degree of the class ck attached to the leaf node lr.
Now, the transformation of this FDT into fuzzy rules is described. In fuzzy cases,
a new instance e may be classified into different classes with different degrees. In other
words, the instance e is classified into each possible class ck with a degree. For that
reason, a separate group of classification rules can be considered for each class ck,
1kmc. Lets suppose the group of classification rules for ck. Then, each leaf lr L
corresponds to one r-th classification rule. The condition of the r-th classification rule
is a group of attribute is attributes value which are connected with operator AND.
These attributes in the condition of r-th rule are attributes associated with the nodes in
the path from the root to the r-th leaf. The attributes values are the values associated
with the respective outgoing branches of the nodes in the path. The conclusion of the rth rule is C is ck. Lets consider that in the path from the root to the r-th leaf there are
q nodes associated with attributes Ai1, Ai2, , Aiq and respectively their q outgoing
branches associated with the values ai1,j1, ai2,j2, ,aiq,jq. Then the r-th rule has the
following form:
IF Ai1 is ai1,j1 AND Ai2 is ai2,j2 AND AND Aiq is aiq,jq THEN C is ck (with
thruthfulness Fkr ).
In the two following paragraphaps, two ways how to use these fuzzy classification
rules for classification of a new instance e into a class ck, 1kmc, are described. The
first way uses only one classification rule whereas the second uses one or more rules
for the classification. The classification means that the degree the instance e is
classified into the class ck with is denoted by membership function ck(e). ck(e) has a
value in the continuous interval 0 to 1. It is also supposed that all attributes values
membership functions of the new instance e, ai,j(e), ai,j Ai, Ai A, are known.
The first way is based on [16] and was initially meant for classification in crisp
cases. In crisp case, an attributes possible value either is the attributes value or isnt.
In other words, either ai,j(e)=1 or ai,j(e)=0 for ai,j Ai, Ai A. Also, just one ai,j(e) is
equal 1 for ai,j Ai, all the others are equal 0. In fuzzy case, there can be more than one
ai,j(e)>0 for ai,j Ai. To be able to use this first way, ai,j(e) with the maximal value
among all ai,j Ai can be set equal 1 and all the others equal 0. The instance with such
membership functions values is called rounded one. Then one rule is chosen from the
made group of rules. It is the r-th rule whose attributes values ai1,,j, ai2,,j, , aiq,,j
respectively match with ai,j with ai,j(e)=1 of the rounded instance. After that, ck(e) is
set equal Fkqr.

Learning Fuzzy Rules from Fuzzy Decision Trees

148

The second way uses one or more classification rules from cks group for
classification of e into a class ck. The reason why it is like this is that because there
may be several ai,j(e)>0 for a node in the FDT. Thats why, there may be several paths
whose all outgoing nodes branches are associated with ai,j where ai,j(e)>0. And each
this path corresponds to one classification rule. But because of the ai,j(e) are not all
equal 1, it is clear that such each rule should be included in the final value of ck(e)
with a certain weight. The weight is for instance e and the r-th rule in cks group given
by the following form:

Wr (e ) =

a path _ r

(e) ,

where a(e) is the value of the membership function of a attributes value a and path_r
is a set of all attributes values in the condition of r-th classification rule. The weight
Wr(e) is equal 0 if there is a attributes value a whose a(e)=0 in the condition of the rth rule. The membership functions value for the new instance e and for class ck is:
R

ck(e)= Fkr *Wr (e) ,

r =1

r
k

where F is the thruthfulness of the r-th rule, or in other words, the certainty degree of
the class ck attached to the leaf node lr.
If, in the first or the second way, classification only into one class is needed,
instance e is classified into the class ck whose ck(e) k=1, 2, , cmc is maximal.
Let us describe the process of transformation of the FDT into fuzzy rules and the
two mentioned ways of their use for classification with the following examples.
Example 1 An instance is described by N=4 attributes A = {A1, A2, A3, A4} = {Outlook,
Temp(erature), Humidity, Wind} and the instances are classified with one class
attribute C. Each attribute Ai = {ai,1, ..., ai,j, , ai,mi} is defined as follows: A1 = {a1,1,
a1,2, a1,3} = {sunny, cloudy, rain}, A2 = {a2,1, a2,2, a2,3} = {hot, mild, cool}, A3 = {a3,1,
a3,2} = {humid, normal}, A4 = {a4,1, a4,2} = {windy, not windy}. The class attribute C =
{c1, ..., ck, , cmc} is defined as follows: C = Plan= {c1, c2, c3} = {volleyball,
swimming, weight lifting}. There is a FDT in Fig.1. The FDT was made from a
database by using cumulative information. The database and the mechanism how to
build it are described in [9]. The FDT has R=9 leaves. In the leaves there are written
values of Fkr for respective ck, k=1, 2, 3. Our goal is to determine ck(e), k=1, 2, 3 for a
new instance e which is described in the Tab.1 on basis of the FDT.

Journal of Information, Control and Management Systems, Vol. 4, (2006), No. 2

149

Temp (A2)
hot (a2,1)

cool (a2,3)
mild (a2,2)

Outlook (A1)

Outlook (A1)
sunny (a1,1)

cloudy (a1,2)

F19=16.3%
F19= 4.9%
F19=78.8%

rain (a1,3)

rain (a1,3)

sunny (a1,1)

F18=30.7%
F15=32.9%
8
5
F2 =23.9% cloudy (a1,2) F2 =10.6%
8
F2 =58.7%
F35=43.2%

F13=37.6% F14=11.0%
F23=50.4% F24=16.3%
F33=12.0% F34=72.7%

Wind (A4)

Wind (A4)
windy (a4,1)
not windy (a4,2)

windy (a4,1)

not windy (a4,2)

F11=14.2%
F21=67.8%
F31=18.0%

F12=33.2%
F22=62.3%
F32= 4.5%

F16=25.1%
F26= 6.7%
F26=68.2%

F17=55.1%
F27=14.2%
F37=30.7%

l1

l2

l6

l7

Figure 1 An example of a FDT adopted from [9].

Table 1 A new instance e with unknown ck(e)=?, k=1, 2, 3.
Attributes
and
their a1,1
values
ai,j(e)
0.9

A1,2

A1
a1,3

a2,1

A2
A2,2

a2,3

a3,1

A3
a3,2

a4,1

A4
A4,2

0.1

0.0

1.0

0.0

0.0

0.8

0.2

0.4

0.6

The classification rules for class c1 (c1group of rules) made from the FDT in Fig.1.
have the following form:
r=1: IF Temp is hot AND Outlook is sunny AND Wind is windy THEN Plan is
volleyball (thruthfulness F11 = 0.142)
r=2: IF Temp is hot AND Outlook is sunny AND Wind is not windy THEN Plan is
volleyball (thruthfulness F12 = 0.332)

Learning Fuzzy Rules from Fuzzy Decision Trees

150

r=3: IF Temp is hot AND Outlook is cloudy THEN Plan is volleyball (thruthfulness F13
= 0.376)

r=9: IF Temp is cool THEN Plan is volleyball (thruthfulness F19 = 0.163)

Similarly, c2group of 9 rules and c3group of 9 rules can be determined.
The following Examples 2, 3 show how to use the fuzzy classification rules.
Example 2 (classification with one fuzzy rule). First, instance e in Tab.1 is transformed
into rounded instance e in Tab.2.
Table 2 The rounded instance e.
Attributes
A1
a1,2 a1,3
a2,1
and their
values
ai,j(e)
1
0
0

A2

A3

A4

a2,2

a1,2

a1,3

a2,1

a2,2

a1,2

a4,2

It means that, for each Ai (i=1, 2, 3, 4), ai,j(e) (j=1, , mi) with the maximal
value among all ai,j Ai is set equal 1 and all the others equal 0. When ai,j with
ai,j(e)=1 are chosen, the following set {a1,1, a2,1, a3,1, a4,2} = {sunny, hot, humid, not
windy} is obtained. The elements of the set matches with attributes values of fuzzy
classification rule r=2 in Example 1. Therefore, c1(e) for class c1 is computed by
c1(e) = F12 = 0.332. Similarly, c2(e) = F22 = 0.623 on bases of c2group of rules, c3(e)
= F32 = 0.045 on bases of c3group of rules. The maximum value has c2(e). Because of
it, if classification only into one class is needed, instance e is classified into class c2.
Example 3 (classification with several fuzzy rules). Paths, whose all outgoing nodes
branches which are associated with ai,j whose ai,j(e)>0, are shown in several bold
branches in Fig.1. In this situation, its needed to use rules which correspond to leaves
l1, l2, l3. The respective classification rules r = 1, 2, 3 for class c1 are mentioned in
Example 1. For the rule r=1, path_r = {a2,1, a1,1, a4,1}. And so, for instance e described
in Tab.1., W1(e) = 0.9*1.0*0.4 = 0.36. Similarly, W2(e) = 0.9*1.0*0.6= 0.54, W3(e) =
0.1*1.0 = 0.1. All the other Wr(e) are equal 0, r=4, 5, , 9.
Then, c1(e) = 0.142*0.36 + 0.332*0.54 + 0.376*0.1 + 0.11*0 + + 0.163*0 = 0.268.
When its done for c1, c2, c2(e) = 0.631 and c3(e) = 0.101 will be obtained. The
maximum value has c2(e). And so, if classification only into one class is needed,
instance e is classified into class c2.

Journal of Information, Control and Management Systems, Vol. 4, (2006), No. 2

151

EXPERIMENTAL RESULTS
The main purpose of our experimental study is to compare the two mentioned
ways of using fuzzy classification rules made from FDT based on cumulative
information for classification. Their classification accuracy is also compared with other
methods. They all have been coded in Java and the experimental results are obtained
on AMD Turion64 1.6 GHz with 1024 MB of RAM.
The experiments have carried out on selected Machine Learning databases [6].
First, if it was needed, the databases were fuzzyfied with an algorithm introduced in
[8]. Then we separated each database into 2 parts. We used the first part (70% from the
database) for building classification models. The second part (30% from initial
database) was used for verification these models. This process of separation and
verification was repeated 100 times to obtain the models error rate for respective
databases. The error rate is calculated as the ratio of the number of misclassification
combinations to the total number of combinations.
The results of our experiments are in Tab.3, where CA-MM denotes FDT-based
fuzzy classification rules algorithm proposed in [22], CA-SP denotes a CA-MMs
modification which is introduced in [3], CI-RM-M denotes algorithm [10] for making
FDT with cumulative information and which uses the second way of fuzzy
classification rules induction from previous section 3, CI-RM-0 uses the first way of
fuzzy classification rules induction from previous section 3, NBC denotes Nave Bayes
Classifier, k-NN denotes k-Nearest Neighbour Classifier. The numbers in round
brackets denote the order of the method for respective database, number 1 denotes the
least and also the best error rate. The last row contains average error rate for all
databases and respective methods.
Table 3 The error rates comparison of respective methods.
Database

BUPA
Ecoli
Glass
Haberman
Iris
Pima
Wine
Average

Error rate for given database and classification method.

CA-MM
CA-SP
CI-RM-M CI-RM-O NBC
k-NN
0.4253(4) 0.4174(2)
0.4205(3) 0.4416(6)
0.3910(1)
0.4253(4)
0.3112(6)
0.3053(5) 0.2022(2)
0.2654(4) 0.1547(1)
0.2046(3)
(3)
(2)
(5)
(6)
0.4294
0.3988
0.4647
0.5394
0.3335(1)
0.4430(4)
0.2615(3) 0.2624(5)
0.2609(2) 0.2453(1)
0.3471(6)
0.2615(3)
(2)
(4)
(1)
(5)
(2)
0.0400
0.04067
0.02956
0.04556
0.05044(6)
0.0400
(2)
(1)
(5)
(3)
0.2460
0.2436
0.2563
0.2483
0.3091(6)
0.2509(4)
0.06094(4) 0.04566(2) 0.06509(5) 0.02697(1) 0.05094(3)
0.08170(6)
(6)
0.2951
0.2526(5) 0.2301(1)
0.2518(4) 0.2431(3)
0.2409(2)

Learning Fuzzy Rules from Fuzzy Decision Trees

152

Figure 2 The normalized average error rates comparison.

In Figure 2, there is a normalized average error rate for respective methods. It is
computed as: the respective average error rate / the highest average error rate * 100. It
can be seen CI-RM-M which uses several classification rules for classification is
almost 10% better than CI-RM-O and its the best of all methods compared here.
5
CONCLUSION
This paper investigated the mechanism of making fuzzy rules from FDTs based on
cumulative information and two ways of their use for classification of a new instance
into a class. One way is a direct fuzzy analogy of crisp case when one classification
rule is used for the new instances classification. The other way is a generalization of
the former. This uses several classification rules and each rule is included in the
classifications result with a certain weight. Both the ways were compared with each
other. The generalized way has almost 10% better results than the other one on selected
databases and also is the best of all compared methods.
ACKNOWLEDGEMENTS
This research was supported by the grants of Scientific Grant Agency of the
Ministry of Education of Slovak Republic and the Slovak Academy of Science (grant
ZU/05 VV_MVTS13). We would like to say thank you Dr. Irina Davidovskaja (from
Belarusian State Economic University) for constructive comments.

Journal of Information, Control and Management Systems, Vol. 4, (2006), No. 2

153

REFERENCES
[1]
[1]
[2]
[3]
[4]
[5]

[6]

[7]
[8]

[9]

[10]

[11]
[12]
[13]
[14]
[15]

SPOSITO, V. A: Linear programming with statistical applications, Iowa, USA

1989.
AI-SHARNAN, S., KARRAY, F., GUAIEB, W., BASIR, O.: Fuzzy Entropy: a
Brief Survey. IEEE International Fuzzy Systems Conference (2001).
BERKA, P.: Knowledge discovery in databases. Academia, Praha, 2003
(Czech).
BOHIK, J.: An Analysis of algorithms for making fuzzy rules. Master
Thesis, The University of ilina, 2006 (Slovak).
DONG, M., KOTHARI, R.: Look-ahead based fuzzy decision tree induction.
IEEE Transactions on Fuzzy Systems 3, Vol. 9 (2001), 461-468.
GAROFALAKIS, M., HYUN, D., RASTOGI, R., SHIM, K.: Building Decision
Trees with Constraints. Data Mining and Knowledge Discovery 7 (2003), 187214.
HETTICH, S., BLAKE, C. L., MERZ ,C. J.: UCI Repository of Machine
Learning Databases [http://www.ics.uci.edu/~mlearn/MLRepository.html].
Irvine, CA, University of California, Dept. of Information and Computer
Science (1998).
JOHN, G. H.: Enhacements to the Data Mining Process. PhD Thesis, Dept. of
Computer Science, Standford University (1997).
LEE, H.-M., CHEN, C.-M., CHEN, J.-M., JOU, Y.-L.: An efficient fuzzy
classifier with feature selection based on fuzzy entropy. IEEE Transaction on
Systems , Man, and Cybernetics Part B: Cybernetics 3, vol. 31 (2001), 426432.
LEVASHENKO, V., KOVALK, ., DAVIDOVSKAJA, I.: Stable fuzzy
decision tree induction. Annual Database Conference DATAKON 2005, Brno
(2005), 273-280.
LEVASHENKO, V., ZAITSEVA, E.: Usage of New Information Estimations
for Induction of Fuzzy Decision Trees. Proc of the 3rd IEEE Int. Conf. on
Intelligent Data Eng. and Automated Learning, Kluwer Publisher (2002), 493499.
MATIAKO, K.: Database Systems. EDIS Print, ilina, 2002 (Slovak).
MITCHELL, T. M.: Machine Learning. McGraw-Hill Companies, USA, 1997.
MITRA, S., ACHARYA, T.: Data mining multimedia, soft computing, and
bioinformatics. Wiley, 2003.
NILSSON, N. J.: Introduction to machine learning. An early draft of a proposed
textbook, [email protected], 1996.
POOCH, W.: Translation of decision tables. Computing Surveys 2, Vol. 6
(1974).

154

Learning Fuzzy Rules from Fuzzy Decision Trees

[16] QUINLAN, J. R.: Induction of decision trees. Machine Learning 1 (1986), 81106.
[17] UMANOL, M., OKAMOTO, H., HATONO, I., TAMURA, H., KAWACHI, F.,
UMEDZU, U., KINOSHITA J.: Fuzzy decision trees by fuzzy ID3 algorithm
and its application to diagnosis system. IEEE Int. Conf. on Fuzzy Systems
(1994), 2113-2118.
[18] WANG, X. Z., CHEN, B., QIAN, G., YE, F.: On the optimization of fuzzy
decision trees. Fuzzy Sets and Systems 2, Vol. 112 (2000), 117-125.
[19] WEBER, R.: Fuzzy ID3: A class of methods for automatic knowledge
acquisition. 2nd Int. Conf. on Fuzzy Logic and Neural Networks (1992), 265268.
[20] WITTEN, I. H., FRANK, E.: Data Mining Practical Machine Learning Tools
and Techniques with Java Implementations. Morgan Kaufman, 1999.
[21] YEN, J.: Fuzzy Logic - A modern perspective. IEEE Transactions on
Knowledge and Data Engineering 1, Vol. 11 (1999), 153-165.
[22] YUAN, Y., SHAW, M. J: Induction of fuzzy decision trees. Fuzzy Sets and
Systems 69 (1995), 125-139.
[23] ZADEH, L. A.: Fuzzy set. Information and Control 8 (1965), 338-353.