See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309014187

Entropy-based Algorithm for Discretization

Article · April 2011

CITATIONS READS

0 1,549

1 author:

Lamia AbedNoor Muhammed

University of Al-Qadisiyah
12 PUBLICATIONS   27 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

localization of iris in eye image View project

techniques of image processing View project

All content following this page was uploaded by Lamia AbedNoor Muhammed on 12 October 2016.

The user has requested enhancement of the downloaded file.

 Entropy-based Algorithm for Discretization
by
Lamia Al-sawwaff
Computer Science Dep.
College of Computer and Mathematic Sciences
University of Al-Qadissiya
e-mail: [email protected]

Abstract
Discretization is a common process used in data mining applications

d
that transforms quantitative data into qualitative data. Different methods

re
have been proposed in order to achieve this process. The stand stone in the
discretization algorithm is to find potential cut-points which split continuous
range values into nominal values. So the discretization methods vary
te
according to how to find these cut-points. Entropy based method is one of
discretization methods however using information entropy measure.
is
In this paper, the aim was to use the entropy-based method in the
discretization with a proposed algorithm. This algorithm attempts to find
suitable cut-points through new concepts. It is based on the entropy
eg

information method with statistical tool through several steps . The practical
work was executed on experimented data that were downloaded from UCI
repository data.
nR

Keywords:- Discretization, Entropy, Split, Merge

1- Introduction
U

Discretization is a common process used in data mining applications that

transforms quantitative data into qualitative data. Discretization is an
essential task of data preprocessing not only because some learning methods
do not handle continuous, but also for other important transformed in a set of
intervals are more cognitively relevant for a human interpretation[1].
Many data mining systems work best with qualitative data, where the data
values are discrete descriptive terms. However, lots of data are quantitative being
represented by a numeric value rather a small number of descriptors. One way to
apply existing qualitative systems to such quantitative data is to transform the
data[2]. The goal of descritization is to find a set of cut points to partition the
range of continuous values into small number of intervals that have good class
coherence, which is usually measured by an evaluation function. In addition to
the maximization of interdependence between the class labels and attribute
values, an ideal discretization method should have a secondary goal to minimize
the number of intervals without significant loss of class attribute mutual
dependence[3].
Different discretization algorithms' are exist, but traditionally they can
be can be divided into two categories:
1- Unsupervised (or class-blind) algorithms discretize attributes without
taking into account respective class labels. The two representative
algorithms are equal-width and equal-frequency discretization.
2- Supervised algorithms discretize attributes by taking into account the
interdependence between class labels and the attribute values. The
representative algorithms are:, Information Entropy Minimization,

d
Statistics-based algorithms, class-attribute interdependency algorithms,

re
and clustering-based algorithms[4].
While in recent years, extended supervised discretization algorithms have
been emerged and known as "semi-supervised" discretization algorithms.
te
These algorithm based on the concept; by reducing the information needed
to execute supervised algorithms i.e. non-parametric semi-supervised
is
discretization method that based on the MODL framework (“Minimal
Optimized Description Length”)[**].
eg

2-Discretization algorithm
Discretization algorithm is the set of steps that are required in order to
transform continuouse values into another expression discrete values. It aims
nR

to find the cut-points. The term”cut-points refers to a real value within the
range of continuous values that divides the range into two interval is less
than or equal to the cutpoint and the other interval is greater than the cut-
point[4].
U

Atypical discretization algorithm broadly consists of four steps:

(1) sorting the continuous values of the feature to be discretized, (2)
evaluating a cut-point for splitting or adjacent intervals for merging, (3)
according to some criterion, splitting or merging intervals of continuous
value, and(4) finally stopping at some point.[3]

3- Entropy Measure and Discretization

Entropy is one of the most commonly used discretization measure in the
literatures. Shannon defines entropy of a sample variable X as:
 H ( X ) = -å p x log p x ……… (1)
x

Where x represents a value of X and px its estimated probability of

occurring. It is the average amount of information per event where
information of an event as:
I ( x) = - log p x ……… (2)

Information is high for lower probable events and low otherwise. Hence,
entropy H is the highest when each event is equi-probable[5].
Discretization methods use entropy measures to evaluate candidate
cutpoints. This means that an entropy-based method will use the class
information entropy of candidate partitions to select boundaries for
discretization. Class information entropy is a measure of purity and it

d
measures the amount of information which would be needed to specify to
which class an instance belongs[3].

re
Definition 1: Let T partition the set S of examples into the subset S1,…..Sn.
Let there be k classes C1,……..Ck and let p(Ci, Sj) be the proportion of
te
examples in Sj that have class i. The class entropy Ent() of a subset Sj is
defined as[7]:
Ent(Sj)=∑p(Ci,Sj) log(p(Ci,Sj)) ……… (3)
is
It has been shown that optimal cut points for entropy minimization must be
between examples of different classes.
eg

The generalize classes to multiple attributes is defined as:

Definition 2: Let A is a set of attributes {A1,……..Am} that can be used in
measured of classes entropy C11, ……. ,Cmi. Where Cmi is the state (i) of
nR

attribute (m). The class entropy of Sj is defined as:

Ent(Sj)= ∑∑p(Cmi,Sj) log(p(Cmi,Sj)) ……… (4)
U

3- Proposed descritization algorithm

The proposed descritization algorithm in this paper is based on using
information entropy methods with iterative steps. Also, the algorithm using
statistical measure that is standard deviation, i.e. this measure support the
selection of cut-point. The proposed algorithm based on multiple stages.
First, primary discretization for continuous attribute can be executed using
equal-width, second step is the splitting for candidate subsets(intervals),
third step is merging for specific intervals. In each step there is some criteria
would be used. The proposed algorithm is shown in figure(1).
 1- Prepare data for processing
2- Sorting the interest continuous data X={x1,……xn}
3- Specified a class attribute C={c1,..cm}
4- Discretize continuous attribute's values in primary discretization
with subsets X={S1,….Sk}
5- Compute information entropy Ent(S) according to class
attribute C
6- Compute standard deviation of Ent
7- do while Ent(Sj)>threshold(s)
split Ent(S) into two subsets
Compute Ent(S) for new subsets
8- Do while no. of interval>threshold(i)

d
- Merge the subsets that can produced minimum total Ent and
standard deviation

re
- Compute Ent(I) for new subsets
te
Figure(1) proposed algorithm
is

4- Experiment and Results

eg

4-1 description of experiment data

The practical work was accomplished using data set that are available in
UCI repository for machine learning researches. The data set is "Hepatitis".
nR

It consists of 19 attributes(categorical, integer, real) with 155 instances.

Attribute Information:
U

1. Class: DIE, LIVE

2. AGE: 10, 20, 30, 40, 50, 60, 70, 80
3. SEX: male, female
4. STEROID: no, yes
5. ANTIVIRALS: no, yes
6. FATIGUE: no, yes
7. MALAISE: no, yes
8. ANOREXIA: no, yes
9. LIVER BIG: no, yes
10. LIVER FIRM: no, yes
11. SPLEEN PALPABLE: no, yes
12. SPIDERS: no, yes
13. ASCITES: no, yes
14. VARICES: no, yes
15. BILIRUBIN: 0.39, 0.80, 1.20, 2.00, 3.00, 4.00
16. ALK PHOSPHATE: 33, 80, 120, 160, 200, 250
17. SGOT: 13, 100, 200, 300, 400, 500,
18. ALBUMIN: 2.1, 3.0, 3.8, 4.5, 5.0, 6.0
19. PROTIME: 10, 20, 30, 40, 50, 60, 70, 80, 90
20. HISTOLOGY: no, yes
4-2 Results
The proposed algorithm was experimented on "Hepatitis" data set. The
interest continuous attribute that to be discretized was "AGE". The
distributed of examples in the data set according to attribute "AGE" is
shown in figure(2).

d
re
te
is
eg

Figure(2) distribution of examples according to "AGE" continuous values

All the nominal attributes in data set were candidate to be class attribute and
experimented in primary discretization step of proposed algorithm. Each
nR

candidate class attribute was discretized with equal-width. The widths of

interval(subsets) were (30,20,15,10,7). The information entropy for each
subset was computed and then total entropy and standard deviation were
U

computed as shown in Table(1). The difference between the candidate

attributes in produced result is obvious. Three attributes have best results
(SEX, ASCITES, VARICES). These attributes would be candidate to the
next step of proposed algorithm that is splitting step.
The resulted data from equal-width partition using (7 width) that concerned
with "SEX" attribute are (9 partitions). Then resulted data would be passed
to next step of descritization that is "splitting". But the condition of splitting
was not occurred according to splitting threshold(s) that was chosed(0.7) so
there is no splitting and the data were passed to the merging step. In the
merging step , the partitions that were merged are 8 and 9, 1 and 2, then 7
merged with 8 and 9. So the result partitions are 6 partitions. Each partition
is addressed by one interval of "SEX" attribute as shown in table(2).
 interval no. Interval range entropy value

1 -26 0.05898433

2 27-33 0.04847932
3 34-40 0.05439528

4 41-47 0.04785283

5 48-54 0.04785283
6 55- 0.05578864

d
Table(2) partitianing attribute AGE into interval according to

re
attribute SEX class attribute value

The resulted data from equal-width partition using (7 width) that concerned
te
with " ASCITES " attribute are (9 partitions). Then resulted data would be
passed to next step of descritization that is "splitting". Only partition (5) was
is
candidate for splitting because its entropy is exceed threshold(s). After
splitting, then the merging step would be executed. The partitions that were
eg

merged are 1 and 2,9 and 10, then 4and 5, 6 and. The results are 6 partitions
as shown in table(3).
nR

interval no. range entropy value

1 -26 0.040434003
2 27-33 0.048479321
U

3 34-41 0.055553853
4 42-47 0.079202133
5 48-54 0.047852827
6 55- 0.055788643
Table(3) partitianing attribute AGE into interval according
to
attribute ASCITES class attribute value

At last attribute "VARICES" was candidate to be the a class attribute and its
concerned data are obtained from equal-width partition using (15 width).
Then resulted data were passed to splitting. Only partition (5) was candidate
for splitting because its entropy is exceed threshold(s). After splitting, then
the merging step would be executed. The partitions that were merged are 1
and 2,9 and 10, then 4and 5, 6 and. The results are 6 partitions as shown in
table(4).

interval no. range entropy value

1 -28 0.028647625
2 29-36 0.069408734
3 37-39 0.044218153
4 40-43 0.036729102
5 44-47 0.064257689

d
6 48-51 0.052859411

re
7 52- 0.029567632
Table(4) partitianing attribute AGE into intervals according to
attribute VARICES class attribute value
te
4-3 Discussion
1-The results that were obtained from primary step discretization reveal the
is
difference between different attributes concerned with computing entropy
values and so this step is useful in order to choose the attribute that
candidate to be class attribute, therefore three with minimum entropy value
eg

would be chosen.
2- The practical work in this paper experiment random variable in
measuring the width of intervals to be produced.
nR

3- Using primary step can be terminated when the entropy value will be
less, or the standard deviation is not improved through using different
width(this case can be notice with VARICES attribute) as shown in table(1).
U

4- Using standard deviation with entropy value to guide the splitting or

merging steps.

5- Conclusion
The execution of proposed algorithm reveled some facts:
1- The trained examples must have good distribution for continuous
attribute that to be discretized in order to produce good result.
2- It is convenient to be considered the standard deviation of entropy
value for the total partitions in order to forbidden the
3- Using primary discretization with "equal width" method is suitable for
decrease the iterations of splitting
 References
[1] Bondu, A. and Boulle, M. and Lemaire, V., Loiseau, S. and Duval, B., '
A Non-parametric Semi-supervised Discretization Method
ICDM '08 Proceedings of the 2008 Eighth IEEE International Conference on
Data Mining, IEEE Computer Society Washington, DC, USA .

[2] Fayyad U., Irani K., Multi- Interval Discretization of Continuous-Valued

Attributes for classification Learning, Proccedings of the thirteenth Int. Joint
conference on Artificial Intelligence, Amsterdam, John Wiley & Sons, 1994,
p. 284-432.
[3]Kotsiantis, S. and Kanellopoulos, D., 2006, “Discretization Techniques:

d
A recent survey”, GESTS International Transactions on Computer Science
and Engineering, vol.32(1), p.p. 47-58.

re
[4] Kurgan, L. A. and Cios, K. J., 2004, “CAIN Discretization Algorithm”,
“IEEE Transactions on Knowledge and Data Engineering”, vol.16, No.2.
te
[5] Liu, H., and others, 2002, “ Discretization: An Enabling Technique”,
“Data Mining and knowledge Discovery”, no.6, p.p.393-423, Kluwer
is
Academic Publishers. Manufactured in Netherlands.
[6] Muhlenbach, F. and Rakotomalala, Ricco, 2005, “Discretization of
eg

Continuous Attributes”, Encyclobedia of Data Warehousing and Mining,

John Wang(Ed) 2005, p.p.397-402.
nR
U

View publication stats