Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Hybrid Machine Learning Algorithm for Arrhythmia

Classification Using Stacking Ensemble, Random
Forest and J.48 Algorithm
Onwuka, Ugochukwu C. Asagba, Prince O.
Dept. of Computer Science Dept. of Computer Science
Ignatius Ajuru Uni. Of Education Iwofe, Nigeria University of Port Harcourt, Nigeria

Abstract:- Arrhythmias also known as dysrhythmia is a The term hybrid machine learning algorithm is employed
heart ailment that arises when electrical signals that when an ensemble of heterogeneous collection of learners are
coordinate the heartbeats do not work appropriately, they involved in contrast to other ensemble models where
are often precursors to a number of heart diseases which homogenous collection of learners are mostly used as is the
may be terminal, and early detection and adequate case of bagging or boosting.
treatment can save life, in this paper we propose a
classification technique that blends two good performing Ensemble learning is a machine learning theory where
machine learning algorithms to enhance the accuracy of two or more learners (machine learning algorithms) are trained
detecting arrhythmia using Electrocardiogram (ECG) or utilized on datasets to solve the same task by extracting
data and Weka machine learning tool, these algorithms several predictions then merged into a single composite
include the J.48 and Random Forest algorithms combined prediction [2] Ensemble algorithms coalesces the decisions of
with an ensemble algorithm called Stacking; For this separate classifiers that composes it, in order to improve the
experiment the MIT-BIH ECG dataset from Kaggle.com final prediction. according to [3] It is the procedure of running
was used to train, test and validate the hybrid algorithm. two or more related but different models and then fusing their
This dataset used classified ECG data into the 5 super class outcomes into a single score or spread with the aim of
of arrhythmia approved by the association for the improving the accuracy of predictive analytics and data
advancement of medical instrumentation (AAMI) to be mining applications.
detectable by equipment and methods, they include
normal sinus (N), fusion beat (F), supraventricular ectopic A. Electrocardiogram (ECG)
beat (SVEB), ventricular ectopic beat (VEB), and The electrical activities of the heart (typically of
unknown beat (Q). the hybrid algorithm “stacked random consisting depolarization and repolarization) is captured by
forest and j.48) outperformed the other individual the electrocardiogram, it facilitates the detection and diagnosis
algorithms, the performance metrics gotten include of heart anomalies by quantifying electrical potentials on the
97.63% accuracy, an approximate sensitivity (recall) and human body surface, generating a record of the electrical
Positive predictivity (precision) value of 0.98, other metrics currents associated with heart muscular activities.
includes a weighted precision recall curve area of 0.97,
receiver operator characteristics area of 0.96 and test time The propagation of electrical signals in the heart are
of 1.66 seconds and finally a model size of 38.2mb which is pattern like, thus it results to electrical currents ensuing on the
suitable for building application for mobile devices. surface of the body and electrical potential on the skin surface;
consequently, this potential is picked up and/or quantified with
Keywords:- Machine Learning, Arrhythmia Classification, the aid of electrode or sensors. The electrical potential
ECG, Random Forest, J.48, Stacking Ensemble. difference between the spaces where the electrodes are placed
on the skin surface, are normally enhanced using an
I. INTRODUCTION operational amplifier with optic isolation. Then, the signal is
then fed to a high-pass filter; after which it is then also
Arrhythmia is an ailment that ensues when electrical submitted to an antialiasing low-pass filter. Finally, the
impulses that controls how the heart beats don't work as processed signal shows in an analogical to digital converter.
required, this makes the heart to beat faster than normal, too The graphical illustration (a plot of voltage (mV) against time)
slow, flutter, fibrillate, or suffer early heartbeat known as of this process is called electrocardiogram (ECG). ECG was
premature contraction. sometimes, arrhythmias are precursors first demonstrated on humans by Augustus Desiré Waller in
to cardiac arrest which could be fatal; The past two decades 1887 [4], since then, the heart’s electrical activities have been
have seen considerable advancements in the diagnosis and recorded, however, the capacity to diagnose the normal
management of supraventricular and ventricular arrhythmias cardiac rhythm and arrhythmias became a routine medical
[1], with digital devices being more available, this paper check-up from 1960s.
proposes a classification model for a more accurate detection
of arrhythmia by using a hybrid machine learning algorithm.

IJISRT21NOV273 www.ijisrt.com 536

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
II. RELATED WORK III. MATERIALS AND METHODS

A number of methods for ensemble of classifiers already A. ECG Dataset

exists in literature, some techniques like bagging and adaBoost The database used in this research is the MIT-BIH
train each classifier with a unique subset of the training data; database available at Kaggle.com uploaded and preprocessed
Dietterich and Bakiri explored this technique in dealing with a by Shayan Fazeli [12]. The MIT-BIH database is often used as
problem that required a huge number of classes, they split the it is the most characteristic database for arrhythmia
number of outputs into different sets, then generated an classification and it has been used for most of the published
ensemble of classifier. In another worked they trained each research. Additionally, MIT-BIH is besides the foremost
classifier in an altered subset of the input features [5]. Waske database obtainable for research purposes and has been refined
and Benediktsson explored the use of ensemble of support continuously along the years [13]. The ECG signals are
vector machines, SVMs for a multi-source land cover sampled to 125Hz and is used for training, testing and
classification problem using a balanced dataset [6]. They validation of the proposed algorithm, table 1, summarizes
ensemble support vector machines, training each SVM with a property the dataset used for our machine learning model.
separate data source, this method improved their result
extensively when compared with the outcome achieved Sample count 109,446 rows
employing a single SVM that was trained using the entire Attributes count 188 columns
dataset. Similarly, the effectiveness of ensemble technique Categories (Class) 5
was demonstrated by Duin and Tax, they carried out vast Sampling Frequency (Hz) 125
experimentations of ensemble learning options with a MIT-BIH Arrhythmia
conclusion that a combination of classifiers trained on diverse Original Data Source
Dataset
feature sets are beneficial, particularly if each classifier offers
a yields good result [7]. [' N: 0, S: 1, V: 2, F: 3,
Classes
Q: 4
Various researchers have implemented the use of Table 1: Dataset Summary
different machine learning techniques as well as ensembles to
classify ECG data, some of which are discussed as follows: The Dataset provides 5 classes each represented by
[8] performed a data mining experiment using 3 popular numerical figures as shown in table 1, Normal Beat (N): 0,
data mining algorithms which are ID3, CART and decision Supraventricular ectopic beat (S): 1, Ventricular ectopic beat
table to build an ensemble prediction model using a large (V): 2, Fusion Beat (F): 3 and Unknown Beat (Q): 4
dataset with 10-fold cross validation methods to measure the
unbiased estimate. They concluded with a highest accuracy of B. Preprocessing
86.43%. Before carrying out the experiment to determine the
[9] implemented a feature enrichment convolution neural performance of the algorithms, the raw data is processed to
network (FE-CNN) classifier to predict 2 class of arrhythmia, make it usable, here the preprocessing done by applying a
they realized the FE-CNN by enriching the ECG signals from nominal to numeric filter, the ECG data last column, which
MIT-BIH arrhythmia database into time-frequency images has the class of the dataset is a numeric field, thus will not give
using discrete short-time Fourier transform. These images are us a useful supervised learning classification result, hence, it
used as inputs for the CNN, there results showed that FE-CNN is necessary to convert the numeric field to a nominal field. To
obtained a positive predictive rate of 90.1%, sensitivity of achieve this in the Weka data mining tool (WEKA version
75.6%, and F1 score of 0.82 for the detection of S beats. 8.3.1), we will utilize the Numeric-to-Nominal filter with the
Sensitivity, positive predictively, and F1 score are 92.8%, “-R last” as an attribute as illustrated on figure 1 and 2.
94.5%, and 0.94, respectively, for V beat detection.

[10] put forward a novel learning scheme that

encapsulates a hybrid evolutionary fuzzy-rough feature
selection model with an adaptive neural network ensemble.
The fuzzy-rough method was setup to deal with uncertainty
and impreciseness of real valued gene expression dataset and
evolutionary search concept optimizes the subset selection
process.

[11] proposed a deep neural network-based (DNN)

method to predict 5 forms of heartbeat. To achieve good
results, they eliminated noise from the ECG signals by
applying a low-pass filter on the two-lead heartbeat segments
with 2 seconds length generated from the filtered signals, and
classified by an adaptive ResNet model. The proposed method
was evaluated on the MIT-BIH Arrhythmia Database with the
patient-specific pattern. The overall accuracy was 98.6%. Fig 1: After applying filter NumericToNominal filter with the
attributes -R last

IJISRT21NOV273 www.ijisrt.com 537

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Random Forest algorithm is explained as follows [14]:

Fig 2: Class Label Statistics

C. Machine Learning Algorithms

This section discusses the classification algorithms
choosing to build the proposed hybrid classification
algorithms, which consists of 3 machine learning algorithms;
these algorithms include one base-learner (Random Forest),
one meta-learner (J48) and one ensemble algorithm (stacking).
These algorithms are explained as follows:

 Random Forest
Random Forest is an ensemble algorithm that is
 J.48
composed of decision trees (also called “forest”), thus it is an
J.48 machine learning algorithm is WEKA data mining
ensemble of decision trees that uses voting ensemble; as
tool open source Java implementation of Quinlan’s C4.5
shown in Figure 3, to perform classification on a new object
algorithm for making pruned or unpruned decision tree; it is
based on some attributes, each tree provides a classification,
an extension of Quinlan’s prior ID3 algorithm. “C4.5 was
therefore we say the tree “votes” for that class. The forest
previously ranked number one data mining algorithm in 2008”
chooses the classification having the most votes (over all the
according to [15]. A given set of training data that is labeled
trees in the forest).
can be used by J.48 to build decision trees using the concept
of information entropy. It employs the fact that every attribute
of the data can be used to make a decision by splitting the data
into smaller subsets. The decision trees generated by J.48 can
then be used for classification of new unknown data.

J.48 algorithm is explained as follows [16]:

Fig 3: Random Forest Algorithm Depicted

Each tree is planted & grown as follows in random forest:

1. If N cases exists in the training set, then sample of N cases
is taken at random but with replacement. This sample will be
the training set for growing the tree.

2. If there are V input variables, a number v<<V is specified

such that at each node, v variables are selected at random out  Stacking
of the V and the best split on these v is used to split the node. Stacking also called blending by some literature entails
The value of v is held constant during the forest growing. training a learning algorithm to combine the predictions of
several other learning algorithms. Firstly, other algorithms are
3. Each tree is grown to the largest extent possible. There is no trained using the dataset presented, then a combiner algorithm
pruning (that is, reduction of the size of trees). is trained to generate a final prediction with every other
predictions of the other algorithms as added inputs. Stacking
ensemble algorithm is elucidated in the following steps:

IJISRT21NOV273 www.ijisrt.com 538

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
D. Algorithm Steps and distance learning) and unsupervised learning (density
estimation).
1: Learn first-level (Base-learner) classifiers based on the
original training data set. IV. ANALYSIS AND RESULTS
base classifiers are learned, based on dataset provided with a
weight distribution; parameters can be tuned to generate WEKA, an open source machine learning software that
distinct base classifiers for homogeneous classifiers; different provides tools for data preprocessing, visualization and data
classification methods and/or sampling methods can be mining was used in this research; Weka, can be downloaded
applied to generate base classifiers for heterogeneous from the link
classifiers. www.cs.waikato.ac.nz/ml/weka/downloading.html. The
computer system we used in this research is the HP EliteBook
2: A new dataset is built based on the output(s) generated by 8460p with Intel core i7 processor, 8Gb of RAM, 1Gb of
the base classifiers. VRAM and Windows 10 OS.
Output(s) or predicted labels of the base classifiers are held as
new features, and the original class labels are kept as the labels A. Performance metrics
in the new dataset generated. Instead of using predicted labels. To make the right comparison, it is expedient to
understand the measures recommended by the association for
3: Learn a second-level classifier based on the newly the advancement of medical instrumentation AAMI for
constructed data set. evaluating methods. These methods include the following:
Sensitivity (Se), Positive predictivity (+P), False positive rate
Any learning method could be applied to learn second- (FPR), and overall accuracy (Acc). Sensitivity and Positive
level classifier. Predictivity are also known in the literature as recall and
precision, respectively. The ECG dataset used was split by 70-
Stacking algorithm is a general framework, as such we 30, 70% will be used for training and the remaining 30% will
can plug in various classifiers and learning approaches to be used for testing and validation of the algorithms.
create the first-level features and transform the data into a
different feature space. B. Accuracy, Model size and Test Time
Accuracy is one the metric for evaluating classification
Stacking ensemble algorithm is described as follows; models. Informally, accuracy is the fraction of predictions our
algorithm got right. Mathematically, accuracy has the
following definition:

(𝑁𝑢𝑚𝑏𝑒𝑟 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦)

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

For non-binary classification (more than two classes),

accuracy is calculated in terms of positives and negatives as
follows:

(𝑇𝑃 + 𝑇𝑁)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁

Where TP = True Positives, FP = False Positives, FN = False

Negatives and TN = True Negatives.

Table 2 shows the accuracy, model sizes and test time of

the different algorithms examined; Figure 5, Figure 6 and
Figure 7 illustrates the test time, accuracy and model sizes of
Stacking algorithm usually yield improved performance the algorithms graphically for easier appreciation and
than most single trained models. It has been productively used understanding.
on both supervised learning tasks (regression, classification

IJISRT21NOV273 www.ijisrt.com 539

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Accuracy (%)

Build Time

Model Size
Test Time
statistics
Kappa

(Sec)

(Sec)

(Kb)
Random Forest 97.23 0.90 175.50 1.95 38,171
J.48 95.98 0.87 194.72 0.09 527
Stacked Random Forest & J.48 97.63 0.92 1789.64 1.66 38,150
Table 2: Accuracy, Build time, Kappa Statistics, Model size and Test Time

raspberry pi) with the trained model; this is because machine

Test Time (Sec) learning models shouldn’t exceed 50mb for mobile
application.
2.50
1.95
2.00 1.66 C. Positive Predictivity (+P)
Positive predictivity (precision) is defined as the sum of
1.50 true positives (TP) over the sum of true positives and false
1.00 positives (FP); it is an indicator of how sure we are of our true
positive results, high scores for precision indicates that the
0.50 0.09 classifier is returning accurate true positives.
0.00
RF j.48 Stacked RF.J48 Mathematically, Positive predictivity is given as
𝑇𝑝
Fig 4: Test time for the various algorithms +𝑃 =
𝑇𝑝 + 𝐹𝑝

Model Accuracy (%) Table 3 shows the Positive predictivity of the

experimented algorithms; it can be seen that Stacked Random
98 97.63 Forest and j48 performed best with weighted average of 0.98.
97.5 97.23

Stacked RF &
97
96.5
J.48

J48
RF
95.98
96
95.5
95 0 (N) 0.973 0.122 0.98
RF j.48 Stacked RF.J48 1 (S) 0.766 0.121 0.887
Fig 5: Model Accuracy 2 (V) 0.886 0.198 0.954
3 (F) 0.726 0.903 0.865
4 (Q) 0.949 0.98 0.987
Algorithm Model Size (Kb) Weighted average 0.958 0.195 0.976
Table 3: Positive predictivity
50000
38171 38150
40000
Positive Predictivity
30000
1.2
0.958 0.976
20000 1

10000 0.8
527
0.6
0
RF j.48 Stacked RF.J48 0.4
0.195
Fig 6: Algorithm model sizes 0.2
0
The results show that the stacked random forest has the RF J.48 SRJ
best accuracy of 97.63% and a model size of 38.15mb, which
is an improved performance. The model size is worthy of note, Fig 7: Weighted Positive predictivity
if one plans to develop for a mobile application (eg android or

IJISRT21NOV273 www.ijisrt.com 540

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
D. Sensitivity (Se) Table 5 shows the FPR of the experimented algorithms,
Sensitivity (Recall) is defined as the number of true Stacked Random Forest and j48 also performed best with an
positives (Tp) over the number of true positives plus the average low false positive rate of 0.08.
number of false negatives (Fn). It is a measure how sure we
are that our model is not omitting any positive value, high
scores for recall indicates that model trained is performing as

Stacked RF &
expected and returning a majority of all positive results.

J.48

J48
RF
𝑇𝑝
𝑆𝑒 =
𝑇𝑝 + 𝐹𝑛

Table 4 shows the recall of experimented algorithms, it 0 (N) 0.13 0.147 0.096
can also be seen that Stacked Random Forest and J48 1 (S) 0.005 0 0.003
performed best with weighted recall average of 0.976 2 (V) 0.008 0.001 0.003
3 (F) 0.002 0 0.001
4 (Q) 0.004 0 0.001
Stacked RF &

Weighted average 0.109 0.122 0.08

J.48

J48
RF

Table 5: False Positive Rate

False Postive Rate (FPR)

0 (N) 0.984 0.999 0.994
0.14
1 (S) 0.627 0.635 0.724
2 (V) 0.862 0.865 0.905 0.12
3 (F) 0.582 0.566 0.689 0.1
4 (Q) 0.939 0.934 0.962 0.08
Weighted 0.96 0.972 0.976 0.06
average 0.04
Table 4: Sensitivity
0.02
0
0.98 Sensitivity RF J.48 SRJ
0.976
0.975 0.972 Fig 9: Weighted Average FPR of the Algorithms
0.97
F. Receiver Operator Characteristics (ROC) of the
0.965 Algorithms
0.96 ROC is a measuring standard in medical and biological
0.96
machine learning algorithms; that helps one evaluate the
0.955 effectiveness of the models created. The ROC curve
0.95
statistically models false positive and false negative detections
in noisy environments, ROC area represents the performance
RF J.48 SRJ
averaged over all possible cost ratios.

The ROC area has various prediction level given as

Fig 8: Weighted Sensitivity (recall) of the Algorithms follows:
1.0 = Perfect, 0.9 = Excellent, 0.8 = Good, 0.7 = Mediocre,
E. False Positive Rate (FPR) 0.6 = Poor, 0.5 = Totally random, < 0.5 = Invalid
FPR is the ratio of samples not belonging to a given (say 2.0
class A) that was inaccurately classified as that class (class A),
the lower FPR a trained model gives the better the model. Shown in table 6 is the ROC of the algorithms, stacked
random forest performed excellently with the ROC value of
𝐹𝑝
𝐹𝑃𝑅 = 0.964
𝐹𝑝 + 𝑇𝑛

IJISRT21NOV273 www.ijisrt.com 541

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Stacked RF &
Precision Recall Curve Area

J.48

J48
RF
1 0.992

0.98 0.973

0 (N) 0.936 0.994 0.965 0.96

1 (S) 0.856 0.985 0.904 0.938
0.94
2 (V) 0.926 0.998 0.968
3 (F) 0.831 0.982 0.913 0.92
4 (Q) 0.973 0.999 0.984
0.9
Weighted average 0.935 0.995 0.964
RF J.48 SRJ
Table 6: Receiver Operator Characteristics Area of the
algorithms Fig 11: Weighted Average PRC of the Algorithms

V. CONCLUSION
ROC
0.995 The experimental results reveals that the hybrid
1
algorithm, stacked random forest and J.48 performed better
0.98 than the individual algorithms on the MIT-BIH arrhythmia
0.964 dataset with a good accuracy of 97.63%, an approximate recall
0.96 and precision value of 0.98, PRC area of 0.97, ROC area of
0.935 0.96 reassures its effectiveness at providing excellent results
0.94
and a test time of 1.66sec. the hybrid algorithm (Stacked
0.92 Random forest and J.48) performed brilliantly thus is a better
choice for automatic arrhythmia application design. Though
0.9 the model size of 38.2mb is a bit large, it is still a good model
RF J.48 SRJ size for machine learning application design for mobile
Fig 10: Weighted Average ROC of the Algorithms devices, given that benchmark size for mobile application
machine learning models is 50mb.
G. Precision Recall Curve Area
A high area under the curve represents both high recall REFERENCES
and high precision, as show in table 7.
[1]. Aro, A. L., & Chugh, S. S. (2018). Epidemiology and
global burden of arrhythmias (Vol. 1).
Stacked RF & J48

https://doi.org/10.1093/med/9780198784906.003.0064
[2]. Onwuka, U. (2019). Ensemble Learning. Retrieved July
J.48

22, 2019, from http://megacomsnet.com.ng/ensemble-

RF

learning
[3]. Burn, E. (2015) What is ensemble modeling? -
TechTarget. Retrieved July 23, 2019, from
0 (N) 0.972 0.999 0.986 https://searchbusinessanalytics.techtarget.com/definitio
1 (S) 0.53 0.863 0.764 n/Ensemble-modeling
2 (V) 0.786 0.98 0.927 [4]. Oxford D (2004). Waller, Augustus Désiré (1856–1922),
physiologist. https://doi.org/10.1093/ref:odnb/38099
3 (F) 0.462 0.825 0.723
[5]. Dietterich, T. G., & Bakiri, G. (1991). Error-correcting
4 (Q) 0.885 0.994 0.968
output codes: A general method for improving multiclass
Weighted average 0.938 0.992 0.973
inductive learning programs. AAAI Press. AAAI. 572–
Table 7: Precision Recall Curve Area 577
[6]. Waske B., Benediktsson, J. A. (2007) Fusion of Support
Vector Machines for classification of multisensor data,
IEEE Trans. Geosci. Remote Sens. 3858–3866.
[7]. Duin, Robert & Tax, David. (2000). Experiments with
Classifier Combining Rules. 16-29. 10.1007/3-540-
45014-9_2.

IJISRT21NOV273 www.ijisrt.com 542

Volume 6, Issue 11, November – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[8]. Chaurasia, Vikas and Pal, Saurabh (2013) Early
Prediction of Heart Diseases Using Data Mining
Techniques Caribbean Journal of Science and
Technology, Vol. 1, 208-217, Available at SSRN:
https://ssrn.com/abstract=2991237
[9]. Xie, Qingsong & Tu, Shikui & Wang, Guoxing & Lian,
Yong & Xu, Lei. (2019). Feature Enrichment Based
Convolutional Neural Network for Heartbeat
Classification from Electrocardiogram. IEEE Access.
PP. 1-1. 10.1109/ACCESS.2019.2948857.
[10]. Dash, Sujata & Patra, Bichitrananda. (2020). Genetic
Diagnosis of Cancer by Evolutionary Fuzzy-Rough
based Neural-Network Ensemble. 10.4018/978-1-7998-
1204-3.ch036.
[11]. Zhao, W., Hu, J., Jia, D., Wang, H., Li, Z., Yan, C., &
You, T. (2019). Deep Learning Based Patient-Specific
Classification of Arrhythmia on ECG signal.
Proceedings of the Annual International Conference of
the IEEE Engineering in Medicine and Biology Society,
EMBS, (July), 1500–1503.
https://doi.org/10.1109/EMBC.2019.8856650
[12]. Kaggle. (2018). ECG Heartbeat Categorization Dataset |
Kaggle. Retrieved July 9, 2020, from
https://www.kaggle.com/shayanfazeli/heartbeat
[13]. Luz, E. J. da S., Schwartz, W. R., Cámara-Chávez, G., &
Menotti, D. (2016). ECG-based heartbeat classification
for arrhythmia detection: A survey. Computer Methods
and Programs in Biomedicine, 127, 144–164.
https://doi.org/10.1016/j.cmpb.2015.12.008
[14]. Maheswari K M. Uma, Ashwin Pranesh, S Govindarajan
(2018). "Network Anomaly Detector using Machine
Learning", International Journal of Engineering &
Technology
[15]. X. Wu, V. Kumar, J.R. Quinlan (2008). Knowledge and
Information Systems, 14(1), 1~37.
[16]. The Free Library. (2014). A comparative study of data
mining algorithms for decision tree approaches using
WEKA tool. - Free Online Library. Retrieved September
10, 2020, from
https://www.thefreelibrary.com/A+comparative+study+
of+data+mining+algorithms+for+decision+tree...-
a0505467547

IJISRT21NOV273 www.ijisrt.com 543