International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________

A Comparative Study of Text Classification Methods: An Experimental

Approach
Rupali P.Patil

R. P. Bhavsar

Department of Computer Science

S.S.V.P.Ss Lk. Dr. P. R. Ghogrey Science College
Dhule, India
[email protected]

School of Computer Sciences

North Maharashtra University
Jalgaon, India
[email protected]

B. V. Pawar
School of Computer Sciences
North Maharashtra University
Jalgaon, India
[email protected]
AbstractText classification is the process in which text document is assigned to one or more predefined categories based on the contents of
document. This paper focuses on experimentation of our implementation of three popular machine learning algorithms and their performance
comparative evaluation on sample English Text document categorization. Three well known classifiers namely Nave Bayes (NB), Centroid
Based (CB) and K-Nearest Neighbor (KNN) were implemented and tested on same dataset R-52 chosen from Reuters-21578 corpus. For
performance evaluation classical metrics like precision, recall and micro and macro F1-measures were used. For statistical comparison of the
three classifiers Randomized Block Design method with T-test was applied. The experimental result exhibited that Centroid based classifier out
performed with 97% Micro F1 measure. NB and KNN also produce satisfactory performance on the test dataset, with 91% Micro F1 measure
and 89% Micro F1 measure respectively.

Keywords-Machine Learning;Nave Bayes; K-Nearest Neighbor; Centroid Based; Text Classification

__________________________________________________*****_________________________________________________
I.

INTRODUCTION

This is an era of computer and Internet. In recent years

numbers of users accessing computers are increased. Because
of development in resources used for communication, data can
be easily sent from one location to another. Thus online
resources are increased. Therefore, today internet is the main
source of information. As a result, information stored on the
web is increasing rapidly. This information may be in the form
of text, numerals, images, graphs, audio and video. It is
continuously growing in size and complexity. One can use this
huge data effectively; if and only if it is properly managed and
organized according to our need. As most of this above
information (above 80%) is stored as text [1], there is a
problem of proper organization and management of this huge
textual information. Classification is helpful in this direction.
Text classification is the process in which text document is
assigned to one or more predefined categories based on the
contents of document. In general, let D be a set of documents,
dj be a document belongs to it and let {c1, c2, , cn} is set of all
classes (text categories), then text classification assign
document dj to a single class or more than one classes. If a
document is assigned to single class it is known as single class
classification where as if it is assigned to more than one classes
then it is known as multi class classification. Text classification
has number of applications such as Email Classification
[2],Topic Spotting on News Articles [3], Language
identification[4] etc. In general, text classification plays an
important role in information extraction and summarization,
text retrieval and question-answering system.

Two major approaches described for text categorization are

Rule based and Machine learning based approaches [5]. In Rule
based approach rules are defined manually and document is
classified according to these rules. This approach is suitable
when document set is small. As rules are defined by human
experts this approach is very accurate. But it totally depends on
rules defined by human experts. Also if the document domain
coverage is large then defining rules can be very tedious job.
Moreover human experts may require writing more rules if
document set increases. In Machine learning based approach
text classifier is built automatically from a set of predefined
classes. That is for construction of classifier, there is no need of
human expert. Thus this approach saves human efforts, time
and till provides comparable accuracy which is achieved by
domain experts.
In machine learning generally two types of learning
algorithms are found in the literature: supervised learning
algorithms and unsupervised learning algorithms [6]. In this
paper we have considered only supervised learning algorithms.
Supervised learning means learning from examples. As
humans learn from past practices, a computer system uses data
to learn, which denote some past practices of an application
area. Goal of supervised learning is to build a classification
model on the basis of the data. The classes of new
cases/instances can be predicted by using this model. In
supervised learning, the data are labeled with pre-defined
classes. In learning (training) phase the training data is used to
learn a model. In testing phase, unseen test data is used to test
the model and to measure the accuracy of model. If training
examples are proper characteristic of the test data, good
accuracy can be achieved on the test data. Many statistical and
517

IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
machine learning techniques have been proposed for document
classification such as Nave Bayes [7], K Nearest Neighbor [8],
Support Vector Machine (SVM) [9], Decision Tree (DT) [10],
Neural Network (NN) [11] etc.
In our study we have considered only supervised learning
methods to learn our classifiers and estimate them on new test
data set. Our study aims to compare three well known
classification algorithms namely NB, KNN and Centroid
Based. The performance of all three classifiers, on same data
set is evaluated and compared by using same performance
evaluation metrics.
The rest of the paper is planned as follows: Section II
summarizes Literature Survey; while section III gives
Methodology and theoretical description of NB, KNN and
Centroid Based text classification algorithms used in this paper.
Section IV describes Experimental setup and trials followed by
Result and Discussion in section V. Section VI gives
conclusion.
II.

LITERATURE SURVEY

Since our aim is to compare NB, Centroid Based and KNN

algorithms for English text we have considered the Reuters
Standard English Dataset. Many works are already done in the
area of text classification for classifying English text [12], [13].
In [14] Taeho Jo addresses two problems (high dimensionality
and sparse distribution) of representing document using
numerical vectors. He proposed a new neural network for text
categorization called NTC (Neural Text Categorizer) which
uses string vectors for document representation rather than
numerical vector. For evaluating traditional (SVM, NB, KNN,
Back Propagation) and proposed approach he has used three
collections Newspage.com, 20NewsGroups and Reuters 21578.
Experimental result shows that NTC is comparable with best
traditional approach back propagation in terms of classification
accuracy and the learning speed. In his study he has shown that
NTC is more practical than others. Susan Dumais, John Platt,
David Heckerman in [15] have compared effectiveness of five
different inductive learning algorithms for text categorization
namely Find Similarity, Nave Bayes, Decision Tree, Bayesian
Networks and Linear SVM, in terms of training speed,
classification speed and classification accuracy. They have also
compared training set size and alternative document
representation. They have used new version of Reuters-21578
collection containing 75% of stories are for training purpose
while remaining 25% are used for testing. For document
representation they used tf*idf weighting for Find Similar
classifier and for other classifiers they used binary
representation. Mutual Information is used for feature selection.
Their experimental results have showed that Support Vector
Machines are more accurate with 92% accuracy on 10 most
frequent categories and 87% accuracy with 118 categories.
SVM is very fast to train and fast to evaluate. While Find
Similar were lowest accurate with 64.6% accuracy for top 10
categories and 61.7% accuracy with all categories. But Find
Similar is fastest learning method in all. In their experiment
they showed that classification accuracy did not improve by
using NLP derived phrases. Vidhya K. A, G. Aghila in [16]
have proposed a hybrid text classification model based on
Rough Set theory and Nave Bayes Classifier. In their proposed
model, for feature reduction Rough set theory is used and for
classification of documents into the predefined categories
Nave Bayes theorem is used by means of the probabilistic
values. The standard dataset Reuters-21578 and 20
Newsgroups are used. Instead of the traditional bag of words

approach, their model maintains a hierarchy of words. The

proposed model improves the classification accuracy by
overcoming the inaccuracy and ambiguity in data set.
BaiRujiang and Liao Junhua in [17] have proposed a hybrid
model RGSC-Rough Set and Genetic Algorithm for SVM
classifier. They have used rough set theory to reduce the feature
vector space and thus improve classification speed. To improve
classification accuracy they present Genetic algorithm
approach for feature selection and parameter optimization.
They compared the proposed RGSC model with KNN and
Decision Tree classifiers. For their experiment they used
Reuters-21578 corpus. In their experiment for RGSC they got
Average precision of 90.7%, Average Recall of 95.1% and
Average F-measure of 92.5% which are greater than average
precision, average recall and average F-measure of KNN and
DT. Their experimental result showed that RGSC method is
more effective than SVM and other traditional method. To
improve the efficiency of basic EM method, Wen Han and
Xiao Nan-feng in [18] have proposed an enhanced EM method,
their approach is semi-supervised classification based on Nave
Bayesian. In their method they first reduce the feature space by
applying DF*ICIF feature selection function and in subsequent
iteration of EM method, using intermediate classifier unlabeled
documents having maximum posterior category probability are
transferred from unlabeled set to labeled collection. In
enhanced EM numbers of iterations are less. Experimental
results demonstrated that enhanced EM method obtains
effective performance in terms of micro average accuracy and
efficiency.
III.

THEORETICAL BACKGROUND OF CLASSIFIERS

This section elaborates the important phases of Classification

process, which by and large includes selection of classification
method, text representation and dimensionality reduction and
learning/training of classifier.
A. Learning of Classifiers
A Classifier is a model which classify new document based
on the previous result of document classification. In this paper
we have compared three supervised learning algorithms for
document classification viz. Nave Bayes, K-Nearest Neighbor
and Centroid Based.
All the three classifiers require some labeled examples to
train the classifiers in training phase, while in testing phase,
unseen data is used to test and evaluate the classifiers. There is
no need to write formal rules to train the classifier rather
general subject knowledge is sufficient. Therefore it is easy to
train the classifier and such inductive classifiers allow users to
give category definition, which is important in some
application.
B. Text Representation and Dimensionality Reduction
Since machine cannot understand the document in its raw
form we have to represent it with some document
representation model. For our case, we have used the popular
vector space model. In our model we have used tf*idf
weighting scheme for KNN and Centroid Based approach
while tf is used for NB classifier. Here tf is term frequency and
idf is inverse document frequency of a term t. Huge
dimensionality is major issue for text classification. Hence we
have to apply dimensionality reduction techniques. We have
used stopword removal and stemming for dimensionality
reduction. Stopword is a meaningless word like a, an, the;
518

IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
whereas stemming is a process of removing suffixes and
prefixes i.e. obtaining the root word (stem). As R-52 dataset
used in our study is already stemmed and stopwords removed
dataset.
C. Selection of Classifiers
1) Nave Bayes: Nave Bayesian is very fast and easy to
implement so it is widely used in many practical systems. It is
well known statistical method whose performance is relatively
good for large datasets, so it is generally used in text
classification problem [19],[20]. It is simple probabilistic
classifier based on Bayes theorem. This classifier makes an
independance assumption i.e. the values of the attributes are
independent [21] given a class of instance this makes it
suitable for large datasets.
Let C = {c1, c2, ,cn} be set of predefined classes and
d={w1,w2,,wm} be a document vector. We have to find
conditional probability P(ci|d) which is the probability of
document d belong to category ci. The document d will be
assigned to category ci which has maximum conditional
probability p(ci|d)

rather than using distances we have used Cosine similarity

measure to find similarity which is calculated using following
formula

Where d1 and d2 are documents vectors.

3) Centroid Based Classifier (CBC): CBC is simple but
effective document classification algorithm. In CBC, we
calculate the centroid vector also called as prototype vector for
each set of documents belonging to same class. If training data
has k classes then total k centroid vectors {
,.., } where
is the centroid of each class i are calculated using following
two methods
a) Arithmetical Average Centroid (AAC): Most commonly
used
initialization
method
for
centroid
based
classifier

where centroid is the arithmetical

average of all document vectors of class ci

b) Cumuli
Geometric
Centroid

(CGC):

where each term will be given a summation

.
(1)
The document is represented by words vector in vector
space model. Therefore it is necessary to separately calculate
the probability of all the words with remaining words and the
resultant probability will get by multiplying them.

The calculation can be simplified by strong assumption of

NB algorithm that values of all the features are independent of
each other in d when document d belongs to category ci then

weight.
Centroid of each category can be used to classify test
document. To classify the test document dt, we have to find
similarity of document vector
with centroid vector of
each category {
, .. , } using cosine similarity finally
assign documents to the class having most similarity value.
That is dt is assign to the class by using
. The advantage of the CB
classification algorithm is that it summarizes the
characteristics of each class, in the form of concept vector. Its
use has been demonstrated for text classification [24].
IV.

This assumption makes NB simple and fast algorithm and

produce good result in most cases. In text classification we
have to calculate p(wj|ci) of each word and p(ci) of each
category. p(d) is constant for all the given categories. Therefore
above equation becomes

and assign the document d to the category with maximum

posterior probability.
2) K-Nearest Neighbor: It is a simple and widely used
classifier [22],[23] for text classification. In training phase
indexing is done and documents are represented in vector
form. In testing phase, distance or similarity of each test
document with each training document is calculated using
distance measure like Euclidean Distance or similarity
measure like Cosine Similarity. Then k-nearest neighbors (k=3
in our case) of test document are determined using these
distances or similarities. Category of the majority of its nearest
neighbors is assigned to the test document. In this paper,

EXPERIMENTAL SETUP AND TRIALS

We have designed an experiment to test the performance of

NB, KNN and CB classifiers. The Experimental Set up is as
given below
A. Experimental Environment:
For our experiment, we have implemented all three above
mentioned classifiers in Java (jdk1.6.0). The experimental trials
were performed on Pentium V with 2GB RAM and results
were evaluated programmatically.
B. Data Set:
We have used Reuters-21578 (R-52) standard dataset
publically available on web link [25], in which stopword
removal and stemming is already performed. Total 52
categories are available for training and testing purpose with
total 6532 training documents and total 2568 testing documents
on 52 categories. Out of these, we have chosen 10 categories.
Following table (TABLE I) summarizes categories and their
distribution in the form of total number of documents.

519
IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
TABLE I.

DISTRIBUTION OF TRAINING AND TESTING DOCUMENTS

Category

Total
documents

Training

Testing

Alum
Coffee

50
112

31
90

19
22

Macro Average precision (P), Recall (R) and F1 measure (F1)

are also depicted.
TABLE II.

Cocoa
Copper
Cpi

61
44
71

46
31
54

15
13
17

Category

Gnp
Gold

73
90

58
70

15
20

Grain
Jobs
Reserves

51
49
49

41
37
37

10
12
12

Total

650

467

155

C. The Performance Measure:

For quantifying result evaluation most commonly used
performance measures like Recall, Precision and Micro and
Macro F1 measure are used in our experiment. For text
classification, Precision is the ratio of correct text documents to
the total predicted text documents. Recall is the ratio of the
correct text documents to the total text documents. It can be
calculated as given below:

PERFORMANCE OF NB, CB AND KNN

Nave Bayes

K-Nearest
Neighbor

Centroid Based

P
(%)

R
(%)

F1
(%)

P
(%)

R
(%)

F1
(%)

P
(%)

R
(%)

F1
(%)

Alum

100

74

85

100

89

94

100

63

77

Coffee

100

87

93

100

93

97

100

80

89

Cocoa

92

100

96

100

100

100

92

100

96

Copper

100

85

92

100

100

100

92

92

92

Cpi

100

82

90

100

94

97

100

88

94

Gnp

68

100

81

83

100

91

75

100

86

Gold

91

100

95

100

100

100

100

95

97

Grain

100

90

95

100

100

100

82

90

86

Jobs

100

92

96

100

92

96

91

83

87

Reserves

80

100

89

86

100

92

67

100

80

91

91

91

97

97

97

89

89

89

93

91

91

97

97

97

90

89

88

Micro
Average
Macro
Average

The same can be represented graphically in Figure 1, Figure 2

and Figure 3 below.

Where Tp: True Positive

Fp: False Positive
Fn: False Negative

F1 is calculated from precision and recall metrics. It is the

harmonic mean of precision and recall and it is given by
F1= (2 * Precision * Recall) / (Precision + Recall)

Figure 1. Performance of NB, CB and KNN in terms of Precision

F1-score can be computed on each individual category and

then averaged over all categories; this is known as Macro
averaging. This can be given in equation form as:

Where A is total number of categories. Each category has

equal weight in Macro-averaged F1-measure.
F1-score calculated globally over all the test documents is
called micro averaging. Each document has equal weight in
Micro-averaged F1-measure. By taking the average of F1measure values for each category i Micro-averaged F1-measure
is obtained

V.

Figure 2. Performance of NB, CB and KNN in terms of Recall

RESULTS AND DISCUSSION

Table II shows the performance of NB, CB and KNN

classifiers for the above chosen 10 categories in terms of
precision, recall and F1 measure metrics. Micro Average and

Figure 3. Performance of NB, CB and KNN in terms of F1Measure

520
IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
In our experimental trials we have obtained Micro Average
Precision of 91%, 97% and 89% for NB, CB and KNN
respectively. While, Macro Average Precision of 93%, 97%
and 90% was obtained for NB, CB and KNN respectively. The
precision of each category for CB is higher than other two
methods. This indicates that the CB method perform usually
high precision. This is shown graphically in Figure 4.

scheme. We have used Randomized Block Design method to

compare the F1 measure obtained by different classifiers.
A. Randomized Block Design [26]:
In Randomized Block Design there is only one primary
factor under consideration in the experiment. Similar test
subjects are grouped into blocks. Each block is tested against
all treatment levels of the primary factor at random order. This
is intended to eliminate possible influence by other extraneous
factor.
In our example we considered classifiers as treatments and
classes as blocks. ANOVA table for Randomized Block Design
by using Table III is as given below:
TABLE III.

ANOVA TABLE USING RANDOMIZED BLOCK DESIGN

Source

Figure 4.

Performance of NB, CB and KNN in terms of Micro and Macro

Average Precision

The Micro Average Recall of NB, CB and KNN are 91%,

97% and 89% respectively. The Macro Average Recall of NB,
CB and KNN are 91%, 97% and 89% respectively. The recall
of each category for CB is higher than other two methods. This
indicates that the CB method perform normally high recall,
which is shown in Figure 5.

DF

Sum of
Square

Mean sum
of square

F-Ratio

Treatment

356.6

178.3

16.59

Table
value
3.55455715

Block

530.7

58.97

5.49

2.45628115

Error

18

193.4

10.74

Total

29

1080.7

(DF- Degree of Freedom)

H0: All classifiers do not differ significantly.

H1: At least one of the values differs from others.
=0.05
Test statistics
= 16.59462254 based on DF1=2,
DF2=18
Where MST Mean sum of square of treatment
MSE- Mean sum of square of Error
p-value :p(F>F0) <0.05 using table (Since (16.59462254 >
3.554557)
Since p-value < =0.05 reject H0.
We used T-test to determine where the differences are:
Critical difference at 0.05 levels is 3.07976
TABLE IV.

Figure 5. Performance of NB, CB and KNN in terms of Micro and Macro

Average Recall

The Micro Average F1 of NB, CB and KNN are 91%, 97%

and 89% respectively. The Macro Average F1 of NB, CB and
KNN are 91%, 97% and 88% respectively. The F1 of each
category for CB is higher than other two methods. This
indicates that the CB method outperform NB and KNN as
shown in Figure 6.

STATISTICAL COMPARISON OF DIFFERENT CLASSIFICATION

ALGORITHMS USING T-TEST. THE VALUES GREATER THAN CRITICAL
DIFFERENCE (3.07976) SHOWS THAT CLASSIFIERS IN ROWS ARE STATISTICALLY
BETTER THAN THE CLASSIFIERS IN COLUMNS.

CB

A more accurate comparison of the different schemes can

be obtained by looking at what extends the performance of a
particular scheme is statistically different from that of another

KNN

5.5

8.3

NB

2.8

From this result we can say that centroid based

classification algorithm outperforms all remaining algorithms,
with NB being second and KNN being the last. Treatment
means difference between NB and KNN i.e. 2.8 in this case,
which is less than critical difference 3.07976. Therefore they do
not differ significantly.
VI.

Figure 6. Performance of NB, CB and KNN in terms of Micro and Macro

Average F1

NB

CONCLUSION:

In this paper, we have reported our study on experimental

evaluation of three well known classifiers NB, KNN and CB,
with statistical significance test for English language text
categorization on R-52 of Reuters-21578 already stemmed and
stopword removed corpus. We have compared the performance
of all the three classifiers. No feature selection was applied in
the reported experimentation. The experimental result shows
that all classification results on three classifiers are acceptable
for sample English Language text. The performance of the
Centroid Based classifier is best amongst all classifiers with
521

IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
97% Micro Average F1 measure. Out of the three classifiers
KNN has obtained lowest performance of 89% Micro Average
F1 measure. Statistical comparison of different classification
algorithms shows that Centroid Based algorithm significantly
outperforms NB and KNN. Though the performance of CB is
best, we have observed that the classification speed of NB is
very fast among all three classifiers. We have also observed
that KNN being lazy learning classifier, it is slowest among all.
ACKNOWLEDGMENT
We hereby acknowledge the financial and administrative
support extended under SAP (DRS-I) scheme, UGC New Delhi
at School of Computer Sciences, NMU, Jalgaon.
REFERENCES
[1] Vishal Gupta and Gurpreet S. Lehal,A Survey of Text
Mining Techniques and Applications, Journal of
Emerging Technologies in Web Intelligence, Vol. 1, No.
1, pp. 60-76, August 2009, doi:10.4304/jetwi.1.1.60-76
[2] IzzatAlsmadi,
IkdamAlhami,
Clustering
and
classification of email contents, Journal of King Saud
University - Computer and Information Sciences,Volume
27,
Issue
1,
pp.
46-57,
January
2015, doi: 10.1016/j.jksuci.2014.03.014.
[3] Taeho C. Jo, Jerry H. Seo, Hyeon Kim, Topic Spotting
on News Articles with Topic Repository by Controlled
Indexing, Chapter, Intelligent Data Engineering and
Automated Learning IDEAL 2000. Data Mining,
Financial Engineering, and Intelligent Agents, Volume
1983 of the series Lecture Notes in Computer Science pp
386-391, 27 May 2002.
[4] HidayetTak , TungaGngr, A high performance
Centroid-Based classification approach for language
identification, Pattern Recognition Letters, Volume 33,
Issue 16, 1 December 2012, pp. 2077-2084,
doi:10.1016/j.patrec.2012.06.01.
[5] Pratiksha Y. Pawar and S. H. Gawande, A Comparative
Study on Different Types of Approaches to Text
Categorization, International Journal of Machine
Learning and Computing, Vol. 2, No. 4, August 2012.
[6] Mohammed.abdul.wajeed,
t.
Adilakshmi,
Text
Classification using Machine Learning, Journal of
Theoretical and Applied Information Technology, 2005 2009 JATIT, ISSN: 2229-73673(1-2), pp. 233-237, 2012.
[7] Rish I., An Empirical Study of the Nave Bayes
Classifier,
www.cc.gatech.edu/~isbell/reading
/papers/Rish.pdf.
[8] Taeho Jo*, (2008), Inverted Index based modified
version of KNN for text categorization, Journal of
Information Processing Systems. Vol. 4. No. 1, 2008.
[9] Gharib T. F., Habib M. B., Fayed Z. T., (2009), Arabic
Text Classification Using Support Vector Machines,
wwwhome.cs.utwente.nl/~badiehm/PDF/ISCA2009.pdf
[10] Su J. and Zhang H.,(2006), A Fast Decision Tree
Learning Algorithm, American Association for Artificial
Intelligence, AAAI'06 Proceedings of the 21st national
conference on Artificial intelligence, Vol. 1, 2006, pp.
500-505, ISBN: 978-1-57735-281-5

[11] Fouzi Harrage, Abdul Malik Salman, Al-Salman,

Mohammed BeMohammed, A Comparative Study of
Neural Networks Architectures on Arabic Text
Categorization using Feature Extraction, Publisher:IEEE,
2010 International Conference on Machine and Web
Intelligence (ICMWI), 3-5 Oct. 2010, pg 102- 107,
ISBN:978-1-4244-8608-3,
DOI10.1109/ICMW.2010.5648051
[12] Ms. K. Umamaheswari, Dr. S. Sumathi, Ms. V. Aparna,
Ms. A. Arthi, Text Classification Using Enhanced Nave
Bayes With Genetic Algorithm, International Journal Of
Computer Applications In Xiu-Li Pang, Engineering,
Technology And Sciences (IJ-CA-ETS) Issn: 0974-3596,
Volume 1, Issue 2, April 09 September 09, Page: 263.
[13] Abhishek Sanwaliya, Kripa Shanker and Subhas C. Misra,
Categorization of News Articles: A Model based on
Discriminative Term Extraction method IEEE, Second
International Conference on Advances in Databases,
Knowledge, and Data Applications, 978-0-7695-39812/10,
IEEE Computer Society, 2010,
DOI
10.1109/DBKDA.2010.18
[14] Taeho Jo, April (2010), NTC (Neural Text Categorizer):
Neural Network for Text Categorization, International
Journal of Information Studies Vol. 2 Issue 2.
[15] Susan
Dumais,
John
Platt,
David
Heckerman,MehranSahami
Inductive
Learning
Algorithms
and
Representations
for
Text
Categorization,robotics.stanford.edu/users/sahami/paper
s-dir/cikm98.pdf
[16] Vidhya.K.A, G.Aghila, Hybrid Text Mining Model for
Document Classification, Computer and Automation
Engineering (ICCAE), 2010 The 2nd International
Conference on (Volume:1 ), 26-28 Feb. 2010, pp.210214, DOI:10.1109/ICCAE.2010.5451965
[17] BaiRujiang and Liao Junhua, A Hybrid Document

Classification based on SVM and Rough Sets,

International e-Conference on Advanced Science
and Technology , pp. 18-23, 7-9 Mar. 2009, IEEE
Computer Society, 978-0-7695-3672-9/09 IEEE DOI
10.1109/AST.2009.14.
[18] Wen Han and Xiao Nan-feng, (2011) "An Enhanced EM
Method of Semi-supervised Classification Based on Naive
Bayesian", Eighth IEEE International Conference on
Fuzzy Systems and Knowledge Discovery (FSKD), Vol.
2, P. 987 991, 26-28 July 2011, DOI:978-1-61284-180-9
[19] Kavi Narayana Murthy, Automatic Categorization of
Telugu News Articles, doi: 202.41.85.68.
[20] Kourdi Mohmed EL, Benasid Amine, Rachidi Tajjeeddine, Automatic Arabic Document Categorization
Based on the Nave Bayes Algorithm, Semitic '04
Proceedings of the Workshop on Computational
Approaches to Arabic Script-based Languages pp. 51-58,
Association for Computational Linguistics Stroudsburg,
PA, USA 2004.
[21] Huan Liu and Lei Yu (2005),"Towards Integrating
Feature Selection Algorithm for Classification and
Clustering"
IEEE
Transactions
on
Knowledge
Engineering Vol. 17 No. 4, April 2005.
522

IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 4 Issue: 3

ISSN: 2321-8169
517 - 523

______________________________________________________________________________________
[22] DR. Riyad Al-Shalabi, Dr. GhassanKanaanManaf H.
Gharaibeh, (2006), Arabic Text Categorization Using
KNN
Algorithm.www.uop.edu.jo/download/research/members
/CSIT2006/.../pg20.pdf
[23] K Raghuveer and KaviNarayana Murthy, Text
Categorization in Indian Languages using Machine
Learning Approaches, IICAI 2007: 1864-1883

[24] Eui-Hong(Sam) Han and George Karypis, CentroidBased Document Classification: Analysis & Experimental
Results, Principles of Data Mining and Knowledge
Discovery, pp. 424-431, 2000.
[25] www.cs.umb.edu/`smimarog/textmining/datasets/SomeTe
xtDatasets.html
[26] http://www.r-tutor.com/elementary-statistics/analysisvariance/randomized-block-design

523
IJRITCC | March 2016, Available @ http://www.ijritcc.org

_______________________________________________________________________________________