International Journal of Computer Science Engineering

and Information Technology Research (IJCSEITR)

ISSN(P): 2249-6831; ISSN(E): 2249-7943
Vol. 7, Issue 2, Apr 2017, 17-24
TJPRC Pvt. Ltd.

AN APPROACH TO EMAIL CATEGORIZATION FOR

TELECOMMUNICATION CORPUS

RAJWANT KAUR1 & GAURAV PATHAK2

1
Research Scholar, CSE, Chandigarh University, Ajitgarh, India
2
Assistant Professor, Department of Applied Science, Chandigarh University, Ajitgarh, India
ABSTRACT

At present, most of the transactions and business is taking place through emails and now it is also necessary for
log-in any site. Due to this a large number of emails are collected in our email account which is hard to read, manage.
That is reason for email categorizing. Classifying those emails into categories is a convenient way for people to read them.
So the main aim of this paper is to solve the problem of email overloading by automatically classifies the e-mail into
different classes based on the content of e-mail. Telecommunication industry dataset is used for the categorization. So this
system classifies the email system into two categories service and finance.

KEYWORDS: Email Mining & Classification

Original Article
Received: Jan 17, 2017; Accepted: Feb 28, 2017; Published: Mar 04, 2017; Paper Id.: IJCSEITRAPR20173

1. INTRODUCTION

With the expansion of networks, emails have become effectual, speedy and most prudent forms of
communication. Electronic mail is the method for sending digital messages from one person to another between
computers via a network. [1] Email remnant the most ubiquitous form of communication because of moderate cost
and massive use of the internet. Emails are pre-eminent for signing any social media site, for shopping online, for
online transaction and for online communication. So the number of email users is continually intensified. Acc. to
radicati groups report, there are currently 2.6 billion [2] email active users and by the end of 2019 its growth will
hike up to over 2.9 billion. But the widening of email accounts is growing slightly faster than the number of email
users for the reason that the users have multiple accounts. The proportion of widening of email accounts is 7% per
year. But with the growth of sending email messaging, there has been also substantial growth in unsolicited mail.
The average number of graymail received per user is fourteen which would exceed to nineteen by the end of 2019.
So there is a requisite of email mining. Email mining is not obligatory for graymail filtrating. Despite that, it is also
imperative for email foldering. Today we send and receive 90 messages per day. For some people, it is usually
more than hundred. Hence users spend a lot of their working time on processing the emails and organize these
emails. At the same time, a large part of email traffic consists of business emails, non-personal emails, and friends
emails. People tussle to distinct crucial messages that urging instant attention. So overloading can be tackled by
two ways by email summarization and other is by automatic categorization. Therefore it is uncomplicated to find
and organize both incoming and existing emails.

So in this paper Section 2 reviews the previous work in email mining. Section 3 explains the algorithms,
techniques and dataset that are used in the previous paper in the tabular form. Section 4 presents the results and

www.tjprc.org [email protected]
18 Rajwant Kaur & Gaurav Pathak

section 5 concludes the paper.

2. LITERATURE SURVEY

Two fast machine learning algorithms [3] that are TF-IDF and Nave Bayes [NB] are implemented for the email
categorization and three categories are made. Both the algorithms are contrasted. NB gives good results than the TF-IDF.

Klimt et al. [4] presented the work on email classification based on relationship data. Experiment is conducted on
enron corpus using the SVM classification algorithm. Hence to bring out the terms of emails, parsing is applied and after
that using the ltc formula, weights is assigned. Assessment is done on the premise of F1. CMU dataset is put in which is
self-created by the author to check the performance of the enron. Results are almost similar.

This technical report [5] presents the email categorization based on the timeline using different supervised
learning algorithms. Two large corpuses that are enron and SRI are used for this task. So in the preliminary processing
step, the folders which have fewer messages are deleted. Wide margin Winnow algorithm takes less running time as
comparative to other algorithms. It is also noticed that wide margin algorithm also outperforms when it is compared to
regular winnow.

Xia et al. [6] categorized the emails into the 15 folders for the trouble free access. So for this task two tournament
methods are proposed, namely Round Robin Tournament (RRT) and Elimination Tournament (ET). Firstly both the
tournament methods are contrasted with n-way classification method, in which tournaments methods gives the higher
accuracy. After that ET and RRT are compared in which RRT performs slightly better than ET.

[7]In this paper different classification algorithms are compared which includes J48 decision tree, NB, NN and
SVM for the spam mail filtering.

Li and his team introduce [8] ME model and follows two phase way to categorize the emails bases on the contents
and properties. Then Li started with preprocessing the mails by filtering the non-character symbol, by resolving the links.
In two phase method first it classify the mails into legitimate and spam and in the second phase emails are categorized into
7 categories. For the comparison ME model is tested with NB, SVM, and KNN. ME model is the best one.

This paper [9] implemented the Evolving Email Clustering Method [EECM] that groups the emails based on
users activities. To examine the grouping accuracy of EECM algorithm Davis Bouldin validity index is used, which are
used for measuring the goodness, quality, validity of the grouping technique. So EECM algorithm is compared with
K-means, Fuzzy over the Enron dataset, in which EECM performs better.

Lu et al. [10] proposes the Semantic Vector Space Model [sVSM] for the purpose of email categorization.
The traditional VSM do not contain the semantic relations, so that why the author proposes sVSM method to remove this
problem. So for creating semantic vector, features are extracted by considering the hepernymy-hyponymy relations
between the synonym sets. To assign the weights of sematic vector tf*iwf*iwf algorithm is used. Three experiments are
designed to evaluate the performance of this method. In the first experiment the traditional VSM and sVSM are compared,
in which proposed method performed well. In the second experiment, proposed method is contrast with Bayesian and KNN
algorithms, in which KNN gives higher results as contrast to others. Third experiment shows that with increasing of email
set, categorizing performance also increases.

Impact Factor (JCC): 7.1293 NAAS Rating: 3.76

An Approach to Email Categorization for Telecommunication Corpus 19

Matwin [11] presented the co-training algorithm to solve the problem of unlabeled data by using 1500 emails on
SVM and Nave Bayes. Experiment is conducted using the weka tool. Firstly pre-processing is done by removing the stop
words and by stemming. Problem is divided into highly balanced, medium balanced and balanced. SVM gives better
results than Nave Bayes. Author also tries to find the reason that why Nave Bayes gives poor results than SVM. So they
experiment with the Nave Bayes by removing the features. We see that SVM also outperforms with very large features
than Nave Bayes.

Yang et al. [12] discusses about the spam mails in the healthcare organization. So to classify the emails into spam
or ham, common characteristics are extracted like no drug effects, disease name from the Trec dataset. Different machine
learning algorithms are applied. So to improve the accuracy, Decision tree and nave Bayes algorithms are combined and it
gives higher accuracy as compared to others and also error rate is low.

Kumar et al. [13] compares the fifteen classification algorithms for the classification of email spam. [14] Soni et
al. proposes an AEMS (Automatic Email Management System) for the handling of emails.

Mishra et al. [15] also worked on spam categorization. So, to carry out this task author uses the different tools to
find out the best one. Weka performs better as compares to Rapid miner and support vector machine.

Tang et al. [16] presents the survey on email mining. Author not only reviewed the single task in the email
mining, rather he presented the five major tasks -namely spam detection, contact analysis, email filing, email visualization
and email network property analysis. He also mentions the related techniques and software tools to mine the email. Future
directions are provided by giving the two examples that are email egocentric network and email monetization.

Many classification algorithms are used for the classification of emails to check whether it is legitimate or
non-legitimate. Author [17] performs this experiment in the real environments to check the performance of these
algorithms. So the author collected the email datasets from the university, company, research institute. Results show that
university gives the higher percentage of spam messages due to various subscription services. Decision tree and SVM
gives the better results.

Alsmadi et al. [18] carried out email categorization on the personal email dataset. SVM, KNN, N-gram methods
are developed to achieve clustering and classification of emails. Classification based on N-Gram is shown to be the best as
text is Bi-language.

Table 1: Various Techniques, Algorithms for Email Mining from the Era 2002 to 2015
Paper Year Dataset Techniques Algorithms
[3] 2002 RT Cf TF-IDF, NB
[4] 2004 Er, CMU Cf SVM
[5] 2004 Er, SRI Cf NB,SVM, ME, WMW
[6] 2005 RT Tm Em, RRT
[7] 2007 RT Cf NN, SVM, NB, J48
[8] 2007 RT Cf ME, KNN, SVM, WMW, NB
[9] 2009 RT Cr EECM, K-means
[10] 2010 20-ng Cf tf*iwf*iwf, sVSM, KNN, NB
[11] 2011 RT Cf CoT, SVM, NB
[12] 2012 TC As, Cf, Cr NB, SVM, J48, Km
[13] 2012 Sb Cf ID3, K-NN, SVM, RF, NB, LDA
[14] 2013 20-ng As, Cr, Cf Ap, non-parametric Km++
[15] 2014 Un, Er, SA Cf RF, Bg, SVM, NB

www.tjprc.org [email protected]
20 Rajwant Kaur & Gaurav Pathak

Table 1: Contd.,
[17] 2015 RT Cf SVM, J48, NB
[18] 2015 RT Cf, Cr SVM, Km, NG

RT-Real time, Er-Enron, ng-newsgoup, TC-Trec Corpus, Sb-Spambase, Un-Usenet, SA-Spam Asian, Cf-
Classification, Cr-Clustering, As-Association, Tm-Tournament, TF-IDF-Term Frequency Inverse Document Frequency,
NB-Nave Bayes, SVM-Support Vector Machine, NN-Neural Network, WMW-Wide Margin Winnow, KNN-k-Nearest
Neighbor, ME-Maximum Entropy, Em-Elimination, RRT-Round Robin tournament, Ap-Apriori, EECM- Evolving Email
Clustering Method, Km-K-mean, CoT-Cotraining algorithm RF- Random Forest, Bg-Bagging, NG-NGrams

4. EXPERIMENT

Corpus

In this paper, a telephonic industrys emails are collected. Common data about the emails dataset is composed
from Google provided for categorize the emails based on their content. There are many other public email corpuses
available like enron, spambase, usenet, SRI etc. But some corpuses are used to classify the spam emails and some are
categorized based on the users. So here we build our own dataset.

Emails Content Pre-Processing

A MIME parser is then used to parse information from those emails to make a dataset that contain one record for
every email with the following information parsed: Email file name, email body, subject

Emails Content Data Mining

An automated tool is to further analyze the content from all emails and measure frequency of words. More than
20,000 words are collected. Stemming is also applied in the term frequency table.

E-Mail Clustering

We obtain entire email as centroid and divide into clusters. After dividing into clusters the content, it will pass
through the knowledge dictionary set for scanning. We obtain the score for each cluster. Finally calculate the distance
between cluster to cluster and cluster to original content.

RESULTS

Table 2
Words Cluser1 Cluster2 Email
refund 0.67 0.76 1.43
connection 0.89 0.9 1.79
waiver 1.1 0.23 1.33
billing 0.65 1 1.65
charge 0 0.35 0.35
network 0.78 0.9 1.68
prepaid 1.1 0.9 2
postpaid 0.98 0.99 1.97
service 0.87 0.92 1.79
complaint 0.87 0.76 1.63
issue 0.34 0.54 0.88

Impact Factor (JCC): 7.1293 NAAS Rating: 3.76

An Approach to Email Categorization for Telecommunication Corpus 21

Figure 1: Shows the IDF Values

Table 3: Shows the Accuracy, Precision, and Recall

PROPERTY RESULTS
True Positive 955
True Negative 10
False Positive 15
False Negative 0
Sensitivity (Recall) 97.94%
Precision (Positive Predictive Value) 98.45%
Result Prevalence 97.50%
Accuracy 96.50%

For the assessment different metrics Precision,

Precision Recall, Accuracy are used.

Figure 2: Graphical Representation of Precision, Recall

Work Flow Diagram

Figure 3: Shows E-Mail Pre-Processing

www.tjprc.org [email protected]
22 Rajwant Kaur & Gaurav Pathak

Figure 4: Content Clustering

CONCLUSIONS

Emails classification in particular utilizes several data mining activities such as: Text parsing, stemming,
classification, clustering. There are many goals or reasons why to cluster or classify emails,
emails, his may include reasons such
as: Spam detection, contact analyses, email categorization.
categorization. Results show that our system works perfectly by categorizing
the email into relevant folders.

REFERENCES

1. http://www.dictionary.com

2. www.radicati.com

3. Yong Park, Email categorization

Yang, Jihoon and Sung-Yong categorizati using fast machine learning algorithms, Print ISBN: 978-3-540-
978
00188-1, pp. 316-323,
323, Springer Berlin Heidelberg, 2002

4. Klimt, Bryan and Yiming Yang, The Enron Corpus: A new Dataset
Dataset for Email Classification Research, " In Machine learning:
ECML, Print ISBN: 978-3-540-23105
23105-9, pp.217-226, Springer Berlin Heidelberg, 2004

5. Bekkerman, Ron Automatic categorization of email into folders: Benchmark experiment

experiment on Enron and SRI corpora, 2004

6. Xia, Yunqing, Wei Liu, and Louise Guthrie. "Email categorization with tournament methods." In Natural Language Processing
and Information Systems, Print ISBN: 978-3-540-26031-8,
978 pp. 150-160.
160. Springer Berlin Heidelberg, 2005

7. Youn, Seongwook and Dennis McLeod. Comparative study for Email Classification, Advances and Innovations in Systems,
Computing Sciences and Software Engineering, Print ISBN: 978-1-4020-6263-6,
978 pp.387-391,
391, Springer Netherlands, 2007.

8. Li, Peifeng, Jinhui

hui Li, and Qiaoming Zhu., An approach to Email Categorization with the ME Model, in FLAIRS
Conference, pp. 229-234, 2007

9. Ayodele, Taiwo, Shikun Zhou and Rinat Khausainov, Evolving email clustering method for email grouping: A machine
learning approach, In Applications of Digital Information and Web Technologies, ICADIWT'09. Second International
Conference on the, E-ISBN: 978-11-4244-4457-1 pp.357-362, IEEE, 2009

10. Lu, Zhao and Jianguo Ding, An efficient semantic VSM based email categorization method, International Conference on
Computer Application and System Modeling (ICCASM 2010),Vol.11, E-ISBN:
E 978-1-4244-7273
7273-6, ISSN :2161-9069, pp. 525-
530, IEEE, 2010

Impact Factor (JCC): 7.1293 NAAS Rating: 3.76

An Approach to Email Categorization for Telecommunication Corpus 23

11. Kiritchenko, Svetlana and Stan Matwin, Email classification with co-training, In Proceedings of the 2011 Conference of the
Center for Advanced Studies on Collaborative Research, pp. 301-312. IBM Corp., 2011.

12. Yang, Weiwen and Linchi Kwok, Comparison Study of Email Classification for Health Organizations, International
Conference on Information Management, Innovation Management and Industrial Engineering,Vol.3, ISSN: 2155-1456 pp.468-
473, IEEE, 2012

13. Kumar, R. Kishore, G. Poonkuzhali, and P. Sudhakar. "Comparative study on email spam classifier using data mining
techniques." In the proceedings of the International Multi Conference of Engineers and Computer Scientists, vol. 1, pp. 14-16,
2012

14. Soni, Gunjan, and C. I. Ezeife. "An automatic Email Management Approach Using Data Mining Techniques." In Data
Warehousing and Knowledge Discovery, print ISBN: 978-3-642-40130-5, Series Volume: 8057, online ISBN: 978-3-642-
40131-2, pp. 260-267, Springer Berlin Heidelberg, 2013.

15. Mishra, Ravishankar, and Ramjeevan Singh Thakur, An efficient approach for Supervised Learning Algorithms using
different data mining tools for spam categorization, In Communication Systems and Network Technologies (CSNT), pp.472-
477, IEEE, 2014.

16. Tang, Guanting, Jian Pei and Wo-Shun Luk, Email mining: Tasks, Common Techniques and tools, Knowledge and
Information Systems, Issue 1, Vol. 41, Print ISSN:0219-1377, pp.1-31, 2014

17. Li, Wenjuan and Weizhi Meng. "An empirical study on email classification using supervised machine learning in real
environments." IEEE International Conference, pp.7438-7443, IEEE, 2015

18. Alsmadi, Izzat and Ikdam Alhami. Clustering and Classification of email contents, Journal of King Saud University-
Computer and Information Sciences, Vol.27, Issue 1, doi:10.1016/j.jksuci.2014.03.014, pp. 46-57, 2015

www.tjprc.org [email protected]