Content uploaded by Saadat Alhashmi
Author content
All content in this area was uploaded by Saadat Alhashmi on Dec 15, 2015
Content may be subject to copyright.
IBIMA Publishing
Communications of the IBIMA
http://www.ibimapublishing.com/journals/CIBIMA/cibima.html
Vol. 2010 (2010), Article ID 73408, 8 pages
DOI:10.5171/2010.734081
Copyright © 2010 Lim Wern Han, Saadat M. Alhashmi . This is an open access article distributed under the
Creative Commons Attribution License unported 3.0, which permits unrestricted use, distribution, and
reproduction in any medium, provided that original work is properly cited. Contact author: Lim Wern Han
e–mail: lim.wern.han@infotech.monash.edu.my
Joint Web-Feature (JFEAT): A Novel Web
Page Classification Framework
Lim Wern Han and Saadat M. Alhashmi
Monash University, Kuala Lumpur, Malaysia
_____________________________________________________________________________________
Abstract
With the increasing amount of web pages over the internet, it has been a major concern to obtain information on the
internet accurately at a reasonable cost with decent performance. A potential solution is through the classification of
web pages into meaningful categories. An effective classification of web pages is of benefit to various applications
such as web mining and search engines. Unlike text documents, the nature of web pages limits the performance of
successful traditional pure-text classification methods. Noises exist in the form of HTML tags, multimedia contents,
dynamic contents and the network structure of web pages which requires a deeper look into effective feature
selection of web pages. Often, these features are filtered out relying on the displayed texts of the web page for
classification. This paper proposed a framework where web page features are taken into consideration during
classification of the web page due to the potential valuable information that might be stored within each of the
features. For this reason, this paper explores the potential of the universal Resource Locator (URL), web page title as
well as the metadata for information to be used in classification with various categories defined by the users. The
framework then explores suitable machine learning algorithms for individual classification of each web feature. The
results would then be used for weighted voting to obtain the classification of that webpage. This approach showed
improvements over pure-text as well as virtual-webpage classification approaches.
Keywords
:
web page classification, feature selection, machine learning
____________________________________________________________________________________________
Introduction and Motivation
The objective of this research is to obtain an effective
and efficient automated web page classification
framework through the use of machine learning
algorithms on selected web feature. The classification
of webpage would help the users in obtaining
information.
The World Wide Web is “the universe of network-
accessible information, an embodiment of human
knowledge” as advocated by Tim Berners-Lee. In a
more technical aspect, it is a network of interlinked
hypertext documents storing information connected
via hyperlinks. This ‘web’ continues to expand day-
by-day and as of July 2009, the web consists of
206,026,787 websites according to Netcraft (2010)
[01] making it inefficient to obtain information. To
help facilitate users in obtaining the information they
need, various tools were introduced – net directories
and search engines as suggested by Chekuriand
Goldwasser (1996) [02].
Among the notable net directories on the web would
be the DMOZ Open Directory Project and Yahoo!
directory. Both constructed manually observed by Qi
and Davison (2009) [03] to classify websites into
their categories which allows the users to search for
the set of desired websites within a specific category.
The effectiveness of such taxonomies or directories
in information retrieval has been demonstrated by
Communications of the IBIMA 2
Dumais and Chen (2000) [04] to a great success.
Accurate classification of web pages allows timely
crawling of web pages which would vary based on
their categorization such as the constant update of a
news web page or blog. The classification of a web
page often tells us its nature and structure which is
highly valuable for effective and efficient data
mining of web pages.
As these directories are manually classified and
maintained for accuracy reasons, it is expensive, time
consuming and would never achieve the complete
coverage of the internet. This paved the way for
machine learning methods in automatic classification
of web pages as suggested by Tsukada et al (2001)
[05]. Shen et al (2007) [06] stated that, unlike
successful text-based documents through machine
learning, the hypertext nature of web pages contain
various noises that at times, would contain more
irrelevant information as opposed to relevant ones.
First, we look into similar past researches conducted
(section 1). We then propose that we extract
information from web features with the lowest
probability source of noises which lies with the URL,
webpage title and the Metadata (section 2). These
features are the core elements of a web page which is
easily identified without being extremely dynamic
and volatile.
In section 3, we present our framework – the Joint-
Feature classification framework and related concepts
which supports the framework. We then conducted
experiments on the framework and analyze the results
(section 4). We look into the possible future works to
further extend this research in section5 and concluded
our findings from this research in section 6.
1. Related Works
Over the years, researchers looked into automatic
web page classification with high accuracy and
efficient performance to replace the current manual
classification of web pages.
Tate and (1996)[09] suggested that different type of
web pages would require different evaluation criteria
with the proposal of the 5 most commonly
encountered web pages – advocacy, business,
information, news and personal web pages. One of
the approaches would be through the use of machine
learning techniques as shown by Tsukadaet al (2001)
[05] which obtained high accuracy in automatic
classification in comparison with those of Yahoo!
Japan. Boet al (2009) [15] compares the SU and
Relief feature selection methods with respect to the
performance of 6 classifiers, which 3 of those is
selected to be explored in our paper. Motivated by
this, we would look into various machine learning
algorithms which would be trained to automatically
classify web pages.
Due to the nature of webpages and noises,Shenet al
(2007)[06] use text summarization techniques via an
ensemble method to remove the noise in web pages
in the attempt to improve the performance of web
page classification. Using this method, they are able
to achieve more than 12% improvement over the pure
text method traditionally used for web pages.
Similarly, Choi and Yao (2005)[07] remove noises in
the form of HTML tags, stop words as well as rare
words. It identifies and uses important information
stored in the HTML structure of the web page such as
title, heading and metadata for classification
weighted using a Structure-oriented Weighting
Technique (SWT). Kan and Thi (2005)[10] explore
the use of significant values on various HTML tags
on each keyword which provided improvement on
accuracy.
Besides information found within the web pages
itself, the importance of the uniform resource
location (URL) feature of a web page in classification
was identified in classification of webpages by Kan
and Thi (2005) [10].Shih and Karger (2004)[11]
found success in an automatic web page classification
system that uses URL and their location on the web
page in a tree-like structure to identify web
advertisements. Thus, we would look into the URL of
web pages in our attempt for an effective web page
classifier for various classes instead of only
advertisements.
Chen and Choi (2008)[16] extracted 31 features from
the web page ranging from URL, texts and various
HTML tags to classify web pages into 5 genres with
high success. As the authors stated however, it
requires considerable manual tuning during the
training phase. As the increase in feature increases
the complexity and the overhead of the classifier,
Indraet al (2008)[08] proposed the use of feature
selected during pre-processing. Currently, we
selected 3 features of a webpage to help classify the
webpage.
Qi and Davison (2009)[3] summarizes the various
concepts used for automatic web page classification
with respect to recent works. The distinct difference
with our work is how we classify each feature
3 Communications of the IBIMA
individually and then joint to give the classification
of the web page.
2. Web page features
Currently, we are interested in 3 distinct feature of a
webpage what we believe stores information which
are of importance in the classification of a web page
through machine learning. In the development
process, each of these features would be processed
individually and provide a classification to the web
page based on their accuracy with the machine
learning algorithm of choice. Jointly, these features
would ultimately decide the classification of the web
page.
2.1. Universal Resource Locator (URL)
The universal resource locator of a web page could
be separated into 2 parts – the host and the path. The
host would be the domain name of the web page
which carries the name of the webpage providing a
hint of the classification of the web page. The
extension for the host as well could possibly help out
on the classification of the web page such as “.gov”
for government websites though the “.com” extension
is widely used even though it was meant for
commercial web pages. Similarly, the path of the web
page stores information about the web page due to
their organization within the website through folders
(for example product folders).
In the development process, the URL would be split
through into its host and path. Individually, these
elements would be parsed via n-gram to obtain
tokens which would be crossed referenced with a
dictionary and weighted. For each subdirectory in the
path, there would be a decay of 10% in weights.
2.2. Title
The web page title is one of the most important
features of a web page due to the information it
conveys about the web page in a short piece of text.
Often, the web page title is not left out from web
pages and is easily found under the <title> tag which
could be extracted and processed through a text based
approach.
2.3. Metadata
The metadata of a web page stores information about
the web page which would not be displayed by the
browser found within Meta tags. In this research, we
look into Meta descriptions and Meta keywords. The
Meta description contains a short description about
the web page that could be processed using full text
methods. The Meta keywords contain the keywords
that provide short and accurate information about the
web page. Large number of keywords would be given
penalty for overextending.
3. Joint-Feature Classification
Through our understanding of the nature of web
pages and identifying the features to be used for our
classification (see Section 2); we now propose an
approach to web page classification called Joint-
Feature classification.
3.1. Categorization/ Genre
The category or genre which web pages would be
classified into is flexible to the need of the user via
machine learning classification algorithms. Thus, the
framework is flexible for any classification classes
that are required by the user.
3.2. Training Sources
The classifier would need to be trained using web
pages obtained manually by the user to suit his/ her
classification from the web such as taxonomies
(DMOZ, Yahoo!) or search engines (Google). It is
these training web pages and their associated class
that would decide the performance of the classifier
with respect to the user’s perception of the web page.
It is subjective to the user’s perception and
interpretation on what category/ genre these training
web pages fall into. Thus, the classifier would be
attuned to that user’s need automatically via machine
learning.
3.3. Preprocessing
Each web page would be preprocessed individually to
extract information from the selected feature of the
web pages (see Section 3). Each feature of the web
page is processed separately to suit their nature. The
processing of each feature would be based on the
same bag of words obtained from a suggested java
dictionary (over 83,700 words) to ensure consistency.
Tokens which do not belong to that bag would not be
accepted to be used during training and classification.
Of course, that bag of words could be edited to meet
the requirements of the users.
The result from the preprocessing of these web pages
would be stored into an ARFF file (to be compatible
with WEKA) for each web page feature to be used
for training or classification.
Communications of the IBIMA 4
3.4. Training
Instead of building a virtual web page with all of the
extracted features and their information, we suggest
that these features are classified separately. Each
having feature having its own suitable classifier
(though it might vary based on the training data) with
the highest accuracy for the feature. Thus, one
classifier for each feature would be trained based on
the training data to build the relationship between the
feature and the web page classification.
3.5. Machine Learning Algorithms
We took the machine learning approach to automatic
web page classification where through machine
learning, our solution would learn and classify web
pages into their appropriate category or genre. There
is a wide range of machine learning algorithms out
there. For our work, we would be using the machine
learning tools found within the Waikato Environment
for Knowledge Analysis (WEKA). Among the
considered machine learning algorithms would be: -
- Naïve Bayes (NB) which applies
probabilistic statistical Bayes’s Theorem in
assigning the most likely class to a given
example described by its feature vector [12]
with the assumption that these features are
independent of each other in the
classification (conditional independent).
- Support Vector Machine (SVM), relatively
new machine learning classification
algorithm based on the non-linear mapping
of input vectors to a high-dimension feature
space [13] where a linear decision surface
(hyperplane) is constructed for classification
(binary). For our work, we would be
splitting multi class problems into multiple
binary classifications using the SMO class in
WEKA.
- C4.5 Tree algorithm by Quinlan being one
of the most popular decision tree learning
algorithm where a tree data structure is
created and used to classify new cases based
on a set of training cases where each case
could be described by a set of attributes
[14].
3.6. Looking as a Whole
Given a particular sentence or body of words, we
would look into this body as a combination of the
words rather than just looking at each of the word in
this body and how would they go in hand with
classifiers. For example given a body text of “selling
a book” we would look at how ‘selling’ and ‘book’ as
a whole help to decide that the body belongs to a
business categorization and not looking at ‘selling’ as
a business classification and ‘book’ as an education
classification thus removing the need for a manually
define dictionary. This helps the classification of the
webpage as words has different meaning based on the
context that they are used in.
This could be trained automatically using the
classifier instead of the users defining their own
dictionary to associate each word to a particular
classification (such as a book as an education word).
Thus, it would require less human effort to build and
maintain the classifying model which is one of the
goals of the problem (automatic web page
classification).
3.7. Joint-Feature
The result of the training would be evaluated (0-10
folds depending on the memory load) for each feature
to obtain the accuracy of each feature in
classification. The result from this evaluation would
be used to assign weights to each feature.
When a web page is to be classified, each feature is
classified individually. The results of these
classifications are then multiplied by the weights
given for each feature and then joined together to
decide the classification of the web page based on
their weights.
4. Experimental Analysis
4.1. Training Data
For our testing, we selected 337 web pages from the
Yahoo! directory spanning over 6 categories found
within the ‘popular websites’ recommendations to be
used as our training data. These web pages were
selected based on both our judgment with regard to
the accompanied explanation texts for each link. The
categorization and the number of selected web pages
are shown in Table 1. The unbalance number of web
pages is due to the varying suitable “popular
websites” for each category.
5 Communications of the IBIMA
Table 1. Training data
Category Number of Web Pages
BusinessEconomy 79
Entertainment 38
Government 43
Health 88
News 40
Sports 49
4.2. Training Evaluation
We recorded 325 web pages that were successfully
parsed and processed for training the classifier. Thus,
these 325 were processed successfully to be used to
train the classifier.
From our various evaluation runs (10-folds) of
smaller data size due to memory constraint, we
concluded that SVM would be the most suitable
machine learning algorithm for this set of test data.
The result on the evaluation done for the training data
using SVM with the evaluation tool from WEKA is
available in Table 2. The accuracy of each feature
would be used as the weight of that feature rounded
to the nearest whole number.
Table 2. Training data evaluation (SVM)
Feature Correctly
Classified
Instances
Weight
URL 93.8462 % 94
Title 77.2308 % 77
Meta
keywords
68.0000 % 68
Meta
descriptions
72.6154 % 73
4.3. Evaluation Data
We selected 75 suitable web pages from the Yahoo!
directory of the same categories under the “popular
websites’ of each category as the training data but are
not found within the training data to test our proposed
solution. The unbalanced numbers for each category
is due to the number of web pages in each category
found within the “popular websites” of each category
within the Yahoo! directory with majority of them
selected web pages to be used in training. Web pages
with technical issues such as dead links, containing
foreign language, being under maintenance or time-
out error are omitted from the evaluation.
4.4. Evaluation Results
The result from our evaluation testing is shown in
Table 3 looking at the actual number of web pages
for each category and the number of correctly/
incorrectly classified web pages for each category.
Table 3. Evaluation data result
Category Total
Number
of Web
Pages
Number of Web
Pages Classified as
Correct Incorrect
BusinessEconomy 20 20 15
Entertainment 11 7 0
Government 7 4 0
Health 20 16 0
News 9 5 1
Sports 8 7 0
Total 75 59 16
We measured the effectiveness of our solution using
precision, recall and the f-measure values for each
category in the context of classification. These values
are least affected by the varying number of web
pages for each category unlike that of accuracy
(where high number of correctly classified web pages
of one class would boost up the accuracy of the
evaluation). High precision denotes smaller
classification error of webpages whereas high recall
denotes smaller leak of classification from positive
web pages. The result is shown as in Table 4 and
Table 5.
Table 4. Effectiveness of solution
Category Precision Recall F-Measure
BusinessEco
nomy
57.1429 % 100 % 72.7273 %
Entertainme
nt
100 % 63.6364 % 77.7778 %
Government 100 % 57.1429 % 72.7273 %
Health 100 % 80 % 88.8889 %
News 83.3333 % 55.5556 % 66.6667 %
Sports 100 % 87.5 % 93.3333 %
Average 90.0794 % 73.9725 % 78.6869 %
Communications of the IBIMA 6
4.5. Discussion
From the evaluation results, it could be seen that the
“Business and Economy” category is of high recall
but low precision. In other words, the business
category is overpowering the other categories. Upon
inspection, we note that this is partly due to the
overlapping of the business category with other
categories such as entertainment business
(entertainment), business organization and trade
bodies (government); health business (health),
business and economy news (news) as well as sport
business (sports) over the diversity of training data
based on the Yahoo! classification. The same could
be said for the news category where a government
website (based on Yahoo!) was classified by our
solution as a news website where the website focuses
on governmental news.
We believe that with better definition of classification
categories or genre, the precision of our solution
could be increased as the amount of overlapping
between categories is reduced. When overlapping
could not been prevented, it is suggested that we
allow overlapping of categories as the classification
of a web page is highly subjective based on the
perspective of the users. If the users would personally
train the solution based on their perspective, the
solution would be highly precise to him or her.
Due to the low precision of the classification for the
“Business and Economy” category, the recall of the
other categories is affected particularly where web
pages of those categories are classified as business
web pages over the ones suggested by the Yahoo!
directory’s evaluators. This is particularly obvious
where the business category is highly dominant over
the government and news categories in overlapping
cases. Another possible explanation could also be
attributed to the high number of training web pages
within the business category which increases the
amount of category overlapping web pages to be
trained as business web pages. The solution to this
would be the clear definition of categories.
The advantage of our solution where each feature is
classified individually and joint by weight is that we
could assign more than one classes to a web page by
allowing the classification which come in second by
weights to define the sub category of a particular web
page. For example, a web page about business news
would be a main-class news web page as well as an
off-class business web page. This way, we could
increase the effectiveness of information retrieval
through our solution.
In the investigation, we perform a comparison of
results with that of related work [05] which also uses
Yahoo! (though Yahoo! Japan may differ a little) top
level categories for the evaluation of their solution
over 5 classes against ours of 6 classes. We selected
[5] to benchmark our work against due to the similar
application of our classifier and real-world source of
testing. The comparison took their best results with
the lowest error rate over 5 classes is shown in Table
5.
From Table 5, it could be seen that our solution on
average outperforms the solution from research [05]
that applies decision tress on “basket analysis”
conducted on web pages.
Table 5. Comparison of results with [05]
Results
[05]
Our
Results
Difference
Highest
Precision
87.20 % 100 % + 22.80
Lowest
Precision
67.60 % 57.14 % - 10.46
Average
Precision
79.36 % 90.08 % + 10.72
Highest
Recall
69.00 % 100 % + 31.00
Lowest
Recall
45.50 % 55.56 % + 10.06
Average
Recall
57.30 % 73.97 % + 16.67
5. Future Work
There is still room for various improvements to the
framework. One of the possible extensions to the
framework would be applying machine learning
classifier again on the classification results of each
web page feature instead of weights obtained from
evaluation of the training classifier. This we believe
would have the potential to increase the performance
of our classifier when dealing with a large set of
classes/ categories/ genres to be classified or when
we look further into other features of web pages.
Besides the web page features mentioned in this
paper, we would explore other possible features that
could provide valuable information and hints on the
classification of the web page. One of the identified
features would be through headings of the web page.
If a feature is deemed to be of low significant, it
would be reflected by its weights and on cases where
2 classifications is close for a web page, it could
provide the small deciding push to the classification
of the web page.
7 Communications of the IBIMA
Alternatively, we could explore the network nature of
the web and how can the classification of a web page
be affected by both its inward and outward links.
This would include hyperlinks and hyperlink texts
which would be easily found via their appropriate
tags.
We would also work on the dimension reduction of
our solution as currently we are having over 80,000
attributes (or words in our bag of words) to reduce
the complexity of our solution and increase the
performance of our solution. The complexity freed up
could be used on other areas to increase the
performance such as increasing web page features or
training web pages.
Furthermore, we would look at the time performance
as well as the complexity of our solution which was
not measured at the moment. Currently, we believe
that our solution is computably feasible within
reasonable range from the parsing of web pages to
the training and classification of the solution.
6. Conclusion
In summary, the proposed framework of automatic
web page classification is highly promising from the
results of our evaluation in comparison with that of
manual classification from editors. With an average
accuracy of 90.08%, the automatic classification of
webpages could be used to build taxonomies
automatically that would help users locate their
required information on the web.
The selected web page features do store valuable
information about the web page in the classification
of web pages. Classification of these features
separately and jointly deciding the classification of
the web page is highly successful as opposed to the
approach of text summarization or buildinga virtual
web page
.
We believe that the research into web page
classification is highly promising with the advent of
the web and web 2.0. Successful classification of web
pages especially through automation is highly
rewarding in various fields especially those of data
mining, search engines and so forth.
7. References
[01] Netcraft (2010), “May 2010 Web Server
Survey”. [Online].Netcraft Ltd [Retrieved May 17,
2010],
http://news.netcraft.com/archives/2010/05/14/may_2
010_web_server_survey.html
[02] Chekuri, C. and Goldwasser, M.H. (1996). “Web
Search Using Automatic Classification”
[Online].Stanford University [Retrieved July 22
2009],
http://theory.stanford.edu/people/wass/publications/
Web_Search/Web_Search.html.
[03] Qi, X. and B. D. Davison (2009). "Web page
classification: Features and algorithms." ACM
Computing Surveys (CSUR) 41(2): Article No.: 12.
[04] Dumais, S. and H. Chen (2000). "Hierarchical
classification of Web content." Annual ACM
Conference on Research and Development in
Information Retrieval in the Proceedings of the 23rd
annual international ACM SIGIR conference on
Research and development in information retrieval:
256 - 263 .
[05] Tsukada, M., T. Washio, et al. (2001).
"Automatic Web-Page Classification by Using
Machine Learning Methods." Lecture Notes In
Computer Science in the Proceedings of the First
Asia-Pacific Conference on Web Intelligence:
Research and Development 2198: 303 - 313
[06] Shen, D., Q. Yang, et al. (2007). "Noise
reduction through summarization for Web-page
classification." Information Processing and
Management: an International Journal 43(6): 1735-
1747
[07] Choi, B. and Yao, Z. (2005). “Web Page
Classification: Features and Algorithms”, Book
chapter in “Foundations and Advances in Data
Mining”, Springer-Verlag.
[08] Indra Devi, M., Rajaram, R., &Selvakuberan, K.
(2008), “Generating beat features for web page
classification”, Webology, 5(1), Article 52. Available
at: http://www.webology.ir/2008/v5n1/a52.html.
[09] Tate, M. and J. Alexander (1996). "Teaching
critical evaluation skills for World Wide Web
resources." Computers in Libraries 16(10): 49 – 55
[10] Kan, M.-Y. and H. O. N. Thi (2005). "Fast
webpage classification using URL features."
Conference on Information and Knowledge
Management in the Proceedings of the 14th ACM
international conference on Information and
knowledge management: 325 - 326
Communications of the IBIMA 8
[11] Shih, L. K. and D. R. Karger (2004). "Using urls
and table layout for web classification tasks."
International World Wide Web Conference in the
Proceedings of the 13th international conference on
World Wide Web: 193 - 202
[12] Rish, I. (2001). “An Empirical Study of the
Naïve bayes Classifier”. In: Proceedings of IJCAI-01
Workshop on Empirical Methods in Artificial
Intelligence
[13] Cortes, C. and V. Vapnik (1995). "Support-
Vector Networks." Machine Learning 20(3): 273 -
297.
[14] Salzberg, S. L. (1994). “Book Review: C4.5:
Programs for Machine Learning by J. Ross Quinlan,
Morgan Kaufmann Publisersinc. 1993”. Kluwer
AvademicPublisers.
[15] Bo, S., S. Qiurui, et al. (2009). "A Study on
Automatic Web Pages Categorization." 1423 - 1427.
[16] Chen, G. and B. Choi (2008). "Web Page Genre
Classification." Symposium on Applied Computing
in the Proceedings of the 2008 ACM symposium on
Applied computing: 2353-2357
