Analysis of Comparison of Fuzzy Knn, C4.5 Algorithm, and Naïve Bayes Classification Method for Diabetes Mellitus Diagnosis

International Journal of Computer Applications Technology and Research
Volume 7–Issue 08, 363-369, 2018, ISSN:-2319–8656
www.ijcat.com 363
Analysis of Comparison of Fuzzy Knn, C4.5 Algorithm,
and Naïve Bayes Classification Method for Diabetes
Mellitus Diagnosis
Putri Elfa Mas`udia
Departement of Electrical
Engineering
State Polytechnic of Malang
Malang, Indonesia
Ridwan Rismanto
Departement of Technology
Information
Malang, Indonesia
Abdullah Mas`ud
Departement of Electrical
Engineering
Malang, Indonesia
Abstract: Early detection of diabetes mellitus (DM) can prevent or inhibit complication. There are several laboratory test that must be
done to detect DM. The result of this laboratory test then converted into data training. Data training used in this study generated from
UCI Pima Database with 6 attributes that were used to classify positive or negative diabetes. There are various classification methods
that are commonly used, and in this study three of them were compared, which were fuzzy KNN, C4.5 algorithm and Naïve Bayes
Classifier (NBC) with one identical case. The objective of this study was to create software to classify DM using tested methods and
compared the three methods based on accuracy, precision, and recall. The results showed that the best method was Fuzzy KNN with
average and maximum accuracy reached 96% and 98%, respectively. In second place, NBC method had respective average and
maximum accuracy of 87.5% and 90%. Lastly, C4.5 algorithm had average and maximum accuracy of 79.5% and 86%, respectively.
Keywords : Fuzzy KNN, C4.5 Algorithm, Naïve Bayes Classifier, Diabetes Mellitus
1. INTRODUCTION
Diabetes mellitus (DM) is a disease marked by high
level of blood sugar caused by impaired insulin secretion,
insulin disruption, or both.DM is a heterogeneous group
marked by increase on glucose level in the blood or
hyperglycemia [1].
There are various classification methods, such as K-
nearest neighbor (KNN), fuzzy KNN (F-KNN), decision
treemethod using C4.5 algorithm, Naïve Bayes classifier
(NBC) method, and many other methods. In previous studies,
one of this methods was used to classify a problem without
analyzing which classification method produce the best result.
Yanita Selly conducted a study in 2013 to compare KNN and
F-KNN methods. The result showed that F-KNN method is
better than KNN method, as accuracy of F-KNN reached 98%
while KNN only had 96% accuracy [12].
The result was then further analyzed in this study, where
F-KNN, decision tree method using C4.5 algorithm, and
Naïve Bayes classifier (NBC) method were compared. The
results of these three methods were analyzed to obtain the best
classification method.
A. Research Objective
1. To apply fuzzy KNN method, decision tree method using
C4.5 algorithm, and Naïve Bayes classifier (NBC)
methodin diagnosing DM.
2. To create a software to compare the three methods based
on accuracy, time, precision, and recall.
B. Related Results from Previous Studies
Yanita Selly dkk compared DM classification using
K-nearest neighbor (KNN) and fuzzyKNN methods.
KNN is a classification method that perform strict
prediction on tested data based on k nearest neighbor.
Meanwhile, F-KNN predicts tested data based on
membership value of tested data in each class, and then
class data with highest membership was selected as
resulting predicted class. The study results showed that
F-KNN method is better than KNN method, as accuracy
of F-KNN reached 98% while KNN only had 96%
accuracy [12].
Other study conducted by Parida Purnana regarding
detection of Type II DM using Naïve Bayesbased on
particle swarm optimization. In the study, particle swarm
optimizationwas used to improve accuracy in detecting
DM. The study result showed that this method had
98.16% accuracy and 0.99 AUC, thus it can be classified
as ‘excellent classification’ [9].
Larissa dkk conducted study regarding classification
of client using C4.5 algorithm as creditingbasis. This
study classify clients of a bank, so that when a problem
occurs, the bank could easily obtain rules from the
resulting decision tree. With decision tree method using
C4.5 algorithm, process of gathering information was
faster and more optimal with larger number of data,
therefore the error in decision making could be
minimized [4].
2. SYSTEM PLANNING
The steps of this research were:
1. Studying literatures regarding fuzzy KNN, C4.5
algorithm, and Naïve Bayes classifier methods.
2. Studying dataset from Indian Pima Diabetes that were
used as trainingdata
3. Designing software to perform classification in
accordance with tested methods.
4. Applying fuzzy KNN, C4.5 algorithm, and Naïve
Bayes classifier methodsto diagnose DM.

www.ijcat.com 364
5. Testing and analyzing the results of each method and
calculating the accuracy.
2.1 Data Preprocessing Method
This study used dataset that were then classified as
training data and testingdata. These data were obtained
from UCI machine learning repository database: Indian
Pima Diabetes in http://archieve.ics.uci.edu. There are
768 clinical data in Indian Pima database but not all
attributes are completely available.
768 clinical data obtained from Indian Pima
Database were preprocessed, which means that
insignificant data was deleted to maximize classification
result. Missing valueor data with incomplete attributes
was treated using rules from [LES-12], since the
classification result is highly influential to training data.
From 8 parameter of data, parameters of TSFT and
INS were deleted since the missing value was very large.
The preprocessed data are displayed in Figure 1.
Figure 1 Preprocessed Data
Used training data had six parameters, which were
hamil, ogtt, diastolik, IMB, DPF and Usia. These parameters
were used in classification process. The value of each
parameter was used to determine diabetes diagnosis, where
value of ‘1’ means positive diabetes and ‘0’ means negative
diabetes
2.2 System Description
This study compared three classification methods,
which were fuzzy KNN, C4.5 algorithm, and Naïve Bayes
classifier. Classified data were generated from Indian
Pimadatabase. This study designed a software for
classification process and the results were used to
determine the best method. Process of the study is
displayed in Figure 2.
Figure 2. Flowchart of the Study
2.3 Classification using Fuzzy KNN Method
In general, classification using fuzzy KNN method was
conducted following these steps:
1. Input normalized training data.
2. Determine k value as initial parameter.
3. Determine weight exponent (m), this study used m = 2.
4. Calculate distance between new record data and each
record training data using Euclidian distance.
5. Calculate membership value of each class, class with the
highest membership value then used to determine new
target.
6. Output was the result of the class with the highest
membership value.
2.4 Classification using Fuzzy KNN Method
In Anyanwu journal, Podgorelec explains that C4.5
algorithm is a development of ID3 algorithm that is used
in generating decision tree. C4.5 algorithmis not limited to
binary number and is able to generate decision tree with
multiple variables. Attributes in C4.5 algorithm generate
one branch for every attribute branch in default [7]. Steps
of classification process using decision tree C4.5 in
general can be expressed as flowchart that is displayed in
Figure 3.
Flowchart of decision tree C4.5 algorithm can be
explained as:
1. Training data was required for classification process
2. Since data hamil, Ogtt, Diastolik, IMB, DPF and
Usia were numerical data, early classification was
done to minimize branch for further selection
3. Training data was required for classification process

www.ijcat.com 365
Figure 3. Flowchart of C4.5 Algorithm
4. Frequency of occurrence of each data in positive
and negative diabetes diagnosis was calculated.
5. Branch was determined by calculating entropy and
gain according to aforementioned formula.
6. If initial branch/root had been determined, then the
second branch was determined by removing
parameter of the obtained branch. This process was
repeated until there was no branch candidate.
7. If there was not any branch candidate, then the
process was finished and decision tree had been
generated.
2.4 Classification Using Naïve Bayes
Han (2006) explains that NBC uses Bayesian
algorithm to calculate total probability. In NBC,
probability of one word will be classified as one
category(posterior probability),and it is based on the
highest previous probability (prior probability).Naïve
Bayes works by calculating the number of occurrence of
specific attribute in particular category.
In Naïve Bayes with non-numerical data,
probability of occurrence of specific category can be
directly calculated, then it is multiplied with every
attribute. However, with numerical data, this cannot be
done as the data is continuous. For numerical data, the
probability is calculated using Gaussian equation.
Flowchart of Naïve Bayes is shown in Figure 4.
Figure 4. Flowchart of Naïve Bayes Classifier
3. RESULT AND DISCUSSION
3.1 Classification Data Using Fuzzy Knn
When fuzzy KNN method was selected as
feature process, then the diagnosis could be conducted in
individual or collected data. Figure 5 shows classification
using individual testing data.
Figure 5.Display of Fuzzy KNN Classification using
Individual Testing Data
There were several features in the classification
using testing data collection: 1) input data, which was for
entering the training data and testing data collection; 2)
process FKNN, which was for entering the k and m
parameter values and for clarification. Display of the
clarification of the testing data collection is shown in
Figure 6.

www.ijcat.com 366
Figure 6 Display of the clarification of the testing
data collection Fuzzy KNN
Feature ‘k’ was used to observed results in
accordance with the number of k and feature
‘membership’ was used to observe the membership value,
while feature ‘accuracy’ was used to calculate the system
accuracy whether the results fit the previous theories.
Display of the system accuracy is shown in Figure 7.
Figure 7 Display of Feature ‘System Accuracy
FKNN’
3.2 Classification Data Using Naïve Bayes
process NBC, which was for clarification process. Display
of the clarification of the testing data collection is shown
in Figure 8.
Figure 9 Input Data in Naïve Bayes
Feature ‘Proses NBC’ was used for clarification, while the
feature ‘accuracy’ was used to calculate the level of
system accuracy. The display of the system accuracy is
shown in Figure 10.
Figure 10 Display of the System Accuracy of Naïve Bayes
3.3 Classification Data Using Decision Tree
C4.5
process C4.5, which was for clarification process. Display
of the clarification of the testing data collection is shown
in Figure 10.
Feature ‘accuracy’ was used to observe the level of
accuracy of the algorithm C4.5 classification by the
system and the results were compared with the previous
theories. The display of the results of the accuracy of the
algorithm C4.5 is shown in Figure 11.
Figure 11 Results of the Accuracy of the Algorithm C4.5
3.4 System Testing Method
The methods for system testing were:
1. Training Data Testing, the test was done with equal
amount of the testing data, which was 50, but with
various training data: 80, 120, 160, 200 data
training.
2. Results of the weight exponent (m) in Fuzzy KNN.
This was because m determined how much the
distance weight between each neighbor to the
membership value.

www.ijcat.com 367
3. Duration test for the clarification process between
fuzzy KNN, C4.5 algorithm and Naïve Bayes
classifier
4. Accuracy test between fuzzy KNN, C4.5
algorithmandNaïve Bayes classifier. This test used
accuracy formula.
5. Precision test amongfuzzy KNN, C4.5 algorithm
and Naïve Bayes classifier
6. Recall test fuzzy KNN, C4.5 algorithm and Naïve
Bayes classifier
3.5 Testing and Analysis Results in FKNN
The results of the test using fuzzy KNN method on
the balanced and unbalanced training data are displayed in
Table 1 and Table 2.
Table 1. Result of Fuzzy KNN Test On Balance Training Data
K
System Accuracy (%)
80
training
data
130
training
data
180
training
data
230
training
data
Average
2 70 84 86 88 82
4 80 86 92 94 88
6 92 92 92 94 92.5
8 90 90 94 94 92
10 90 96 96 96 94.5
12 92 96 98 98 96
Table 2. Result of Fuzzy KNN Test On Unbalance Training
Data
K
System Accuracy (%)
80
training
data
130
training
data
180
training
data
230
training
data
Average
2 76 82 82 86 81.5
4 84 86 86 94 87.5
6 84 86 86 94 87.5
8 84 94 96 96 92.5
10 88 94 94 96 93
12 92 96 98 98 96
The results shows that the more training data, the higher
system accuracy. This means that as the number of training
data increases, the number of record with distance near the
predicted data class also increases, which in turn improves the
accuracy.
The test results of balanced training data show that
accuracy tended to increase, except for k=8 where it slightly
decreased. Meanwhile, for 180 and 230 training data, all
system accuracy increased from k=2 to k=12.
Test results of unbalanced training data show that for 80,
130, and 230 training data, all system accuracy increased from
k=2 to k=12. Meanwhile for 180 training data, system
accuracy slightly decreased on k=10.
Test results of both balanced and unbalanced training data
show that the number of training data is directly proportional
to system accuracy. The slight decrease in several tests was
insignificant and system accuracy was tended to be stable.
3.6 Testing and Analysis Results in C4.5
This test was aimed to observe which type of training data
generated the best results. Each training data was tested five
times to obtain the best results.
The test results of balanced training data shows that the best
data was obtained from the second experiment with respective
average and maximum accuracy of 78% and 84% for 130
training data. Meanwhile, the test results of unbalanced
training data shows that the best data was obtained from the
third experiment with respective average and maximum
accuracy of 79.5% and 86% for 80 training data. All test
results are displayed in Table 3
Table 3a. Test Result of C4.5 Algorithm Test on Balance
Training Data
Training
Data
System Accuracy (%)
Balanced Training Data
1 2 3 4 5 Avg
80 70 80 80 88 80 79.6
130 72 84 78 78 80 78.4
180 74 74 74 68 62 70.4
230 74 74 74 74 74 74
Avg 72.5 78 76.5 77 74 75.6
Table 3b. Test Result of C4.5 Algorithm Test on Unbalance
Training Data
Training
Data
System Accuracy (%)
Unbalanced Training Data
1 2 3 4 5 Avg
80 80 86 86 80 82 82.8
130 82 84 84 80 80 82
180 62 62 74 74 74 69.2
230 74 74 74 74 74 74
Avg 74.5 76.5 79.5 77 77.5 77
Each training data was tested five times to observe the
results change, then average of system accuracy was taken.
The results showed that the number of training data is
inversely proportional to system accuracy. This probably
caused by the increasing number of training data makes it
more difficult to generate decision tree.
The test of this study was done on two type training data.
Balanced training data means that the diagnosis was evenly
distributed on positive and negative results. Meanwhile, in
unbalanced training data, the diagnosis was random, which
means that there was no record of the number of positive and
negative results. In this test, average of system accuracy from
two types of training data was calculated. The average results
are shown in Table 4.
Table 4. Average Accuracy of Training Data Type
Number of
Training
Data
Balanced
Training
Data (%)
Unbalanced
TrainingData (%)
80 79.6 82.8
130 78.4 82
180 70.4 69.2
230 74 74
Average 75.6 77

www.ijcat.com 368
The results show that unbalanced training data had better
accuracy with 77% compared with balanced training data
(75.6%).
3. 7 Testing and Analysis Results in NBC
This test was aimed to observe which type of training data
generated the best results. Each training data was tested five
times to obtain the best results. The results can be seen at
Table 5.
Table 5a. Test Result of Classification Using Naïve Bayes
Training
Data
System Accuracy (%)
Balanced Training Data
1 2 3 4 5 Average
80 90 80 86 84 80 84
130 86 88 88 84 82 85.6
180 86 86 90 82 86 86
230 86 86 86 86 86 86
Average 87 85 87.5 84 83.5 85.4
Table 5b. Test Result of Classification Using Naïve Bayes
Training
Data
System Accuracy (%)
Unbalanced Training Data
1 2 3 4 5 Average
80 84 86 80 86 86 84.4
130 84 84 84 88 86 85.7
180 86 86 86 90 84 86.4
230 86 86 86 86 86 86
Average 85 85.5 84 87.5 85.5 85.5
The test results of balanced training data shows that the
best data was obtained from the third experiment with
respective average and maximum accuracy of 87.5% and 90%
for 180 training data. Meanwhile, the test results of
unbalanced training data shows that the best data was
obtained from the fourth experiment with respective average
and maximum accuracy of 87.5% and 90% for 180 training
data. The results show that the number of training data is
directly proportional to system accuracy.
The test of this study was done on two type training data.
Balanced training data means that the diagnosis was evenly
distributed on positive and negative results. Meanwhile, in
unbalanced training data, the diagnosis was random, which
means that there was no record of the number of positive and
negative results. In this test, average of system accuracy from
two types of training data was calculated. The average results
are shown in Table 6.
Table 6. Average Accuracy of Training Data Type
Number of
Training
Data
Balanced
Training Data
(%)
Unbalanced
Training Data (%)
80 84 84.4
130 85.6 85.7
180 86 86.4
230 86 86
Average 85.4 85.5
As can be seen from Table 6, there was no significant
difference between both data types with the accuracy
difference only 0.01%, with unbalanced training data had
slightly higher accuracy than balanced training data.
3. 8 Testing and Analysis Results in Naïve
Bayes
In the previous tests, accuracy of fuzzy KNN, C4.5
algorithm, and Naïve Bayes had been tested in detail. From
the test result, the best accuracy of each method was
compared with other methods without considering training
data type. The comparison result of accuracy of all methods is
displayed in Table 7.
Table 7. Comparison of Accuracy of The Three Methods
Training
Data
Accuracy (%)
Fuzzy
KNN
C4.5 Algorithm
Naïve
Bayes
80 92 86 86
130 96 84 88
180 98 74 90
230 98 74 86
Average 96 79.5 87.5
The result shows that fuzzy KNN method had the highest
accuracy with 96%.
4. CONCLUSSION
From the results, it can be concluded that:
1. The system was able to classify DM diagnosis using fuzzy
KNN, C4.5 algorithm, or Naïve Bayes with average
accuracy of all methods was 88.5%.
2. In classification using fuzzy KNN method, the highest
accuracy was obtained in 180 training data and k=12, with
accuracy of 98% and average accuracy of all training data
of 96%. The results show that the more training data, the
higher system accuracy. This means that as the number of
training data increases, the number of record with distance
near the predicted data class also increases, which in turn
improves the accuracy.
3. In classification using C4.5 algorithm, the highest
accuracy was obtained in 80 training data, with accuracy
of 86% and average accuracy of all training data of
79.5%.
4. In classification using Naïve Bayes method, the highest
accuracy was obtained in 180 training data, with accuracy
of 90% and average accuracy of all training data of
87.5%. The results show that the more training data, the
higher system accuracy.
5. Based on the results of accuracy test, the best
classification method was fuzzy KNN with average
accuracy of 96%, followed by Naïve Bayes method with
87.5%, and lastly C4.5 algorithm with average accuracy
of 79.5%.
6. Based on the results of precision and recall test, the best
classification method was fuzzy KNNwith precision and
recall of 0.94 and 1, respectively. This result shows that
the accuracy is directly proportional to precision and
recall.

www.ijcat.com 369
5. REFERENCES
[1] Brunner and Suddarth. 2002. Buku Ajar Keperawatan
Medikal Bedah, edisi 8 volume 2. Jakarta : EGC.
[2] Keller, James. 1985. A Fuzzy K-Nearest Neighbor. IEEE
vol. SMC-15, No. 4
[3] Kusrini, 2007. Design and implementation of building
decision tree using C4.5 algorithm. Proceedings of
SEAMS-GMU Conference 2007.
[4] Larrisa Navia Rani, 2015. Klasifikasi Nasabah
Menggunakan Algoritma C4.5 Sebagai Dasar
Pemberian Kredit. Jurnal Kom Tek Info Fakultas Ilmu
Komputer Volume 2 No. 2. ISSN : 2356-0010
[5] Li D, Deogun JS, Wang K (2007) Gene Function
Classification Using Fuzzy K-Nearest Neighbor
Approach.
[6] Manning, D. Cristopher, Prabakhar Raghavan dan
Hinrich Schutze. 2009. An Introduction to Information
Retrieval. Cambridge University Press.
[7] Maimon, O. dan Last, M. 2000. Knowledge Discovery
and Data Mining, The Info-Fuzzy Network (IFN)
Methodology. Dordrecht: Kluwer Academic.
[8] Mistra. 2005. 3 Jurus Melawan Diabetes Mellitus.
Jakarta : Puspa Swara.
[9] Purnana, Parida. Dan Supriyatno, Catur.2013 Deteksi
Penyakit DiabetesType II denganNaive Bayes Berbasis
Particle Swarm Optimization. Jurnal Teknologi
Informasi Volume 9 No.2 ISSN 1414-9999.
[10] Sunjana, 2010, Aplikasi Mining Data Mahasiswa
dengan Metode Klasifikasi Decision Tree, Seminar
Nasional Aplikasi Teknologi Informasi, Vol 7 pp. 24-
29.
[11] Tandra, Hans. 2009. Osteoporosis Mengenal,
Mengatasi, dan Mencegah Tulang Keropos. Jakarta:
Gramedia Pustaka Utama.
[12] Yanita, Selly, ridho, ahmad. & lailil. 2013.
Perbandingan K-Nearest Neighbor dan Fuzzy K-
Nearest Neighbor pada Diagnosis Penyakit Diabetes
Melitus.Jurnal Doro Volume 2 no.10
[13] Han, J. and Kamber, M., 2006. Data Mining: Concepts
and Techniques, University of Illinois at Urbana-
Champaign

Analysis of Comparison of Fuzzy Knn, C4.5 Algorithm, and Naïve Bayes Classification Method for Diabetes Mellitus Diagnosis

More Related Content

Similar to Analysis of Comparison of Fuzzy Knn, C4.5 Algorithm, and Naïve Bayes Classification Method for Diabetes Mellitus Diagnosis

More from Editor IJCATR

Recently uploaded

Analysis of Comparison of Fuzzy Knn, C4.5 Algorithm, and Naïve Bayes Classification Method for Diabetes Mellitus Diagnosis