BLOG CLASSIFICATION USING K-MEANS
Ki Jun Lee, Myungjin Lee and Wooju Kim
Yonsei University, seodaemoon-gu seongsan-ro 262, Seoul, Republic of Korea
Keywords: Search, Blog, Classification, Grouping, Clustering, K-means.
Abstract: With the recent exponential growth of blogs, a vast amount of important data has appeared on blogs.
However, dynamic, autonomous, and personal features of such blogs make blog pages be quite different
from those on general web pages in many aspects. As a result, this also causes many problems which cannot
be handled properly by general search engines. One of the problems which we focused in this study is that
blog pages are inherently poorly-organized and very much duplicated. This means the blog search engines
cannot but provide the poorly-organized and duplicated results. To solve this problem, we propose a blog
classification method using K-means and present a blog search result reorganization approach based on this
method. In this study, firstly, we review the current status and their performances of blogs and blog search
engines. Secondly, we adopt the K-means algorithm as a base algorithm and devise a blog title classification
method to reorganize the blog titles resulted by a search engine. Finally, by implementing a prototype
system of our algorithm, we evaluate our algorithm’s effectiveness, and present a conclusion and the
directions for future work. We expect this algorithm can improve the current blog search engines’ usability.
1 INTRODUCTION
As the number of blogs – a compound word
combining ‘web’ and ‘log’ – on the World Wide
Web is growing (Technorati Weblog, 2006), they
have created a brand new subset of the World Wide
Web (Kumar, Novak, Raghavan, and Tomkins,
2003). Because blogs are created based on personal
needs and are updated by individuals, blog posts are
displayed with personal contents like diaries, one’s
experiences, observations, discussions, music,
movies, and other topics, in essence whatever the
bloggers want. This feature of blogs is absolutely
different from those of other web pages, and it has
fostered new research opportunities like information
retrieval, text mining, social studies and blog
searching - which we present in this paper. If some
researcher offers these kinds of blog information
with a more proper result format and more
meaningful classification, it can give users useful
and special information different from the ones
presented by normal web pages.
Since general web search engines do not provide
such a special interface for blogs, there are several
blog search engines only for blogs (Bloglines,
Blogpulse, BLOGRANGER, and BlogWatcher).
Each of these search engines provide their own
interface format based on their own algorithm, such
as blog stats, topic selection, blogger selection, link
selection, and so on. However, this area still has a
long way to go in terms of effectiveness, since no
result from any search engine can completely satisfy
the needs of a variety of users.
There are some examples of the failure to get the
right information when users try to search with a
particular propose in mind. For example, let us
suppose a user is trying to find information about a
novelist the user knows just the name of. When
inputting the novelist’s name as a keyword, the user
may want to know the ranking of the novelist’s
novel, which novel he or she should read first, how
bloggers appreciate the novelist – the novelist’s
reputation - and a book review of the novel by the
average reader, not by a professional book reviewer.
However, it takes a lot of time to satisfy all these
user needs through the general search engine
classification system because of each particular
blog’s features (Aixin, Maggy, and Ying, 2007).
Although some blog search engines support different
categories for blog classification, since blog
information reflects tremendous personal interest,
such a static category has its limitations.
In this paper, we present a blog classification
algorithm using K-means, in an attempt to solve the
61
Jun Lee K., Lee M. and Kim W. (2009).
BLOG CLASSIFICATION USING K-MEANS.
In Proceedings of the 11th International Conference on Enterprise Information Systems - Software Agents and Internet Computing, pages 61-67
DOI: 10.5220/0001949600610067
Copyright
c
SciTePress
problems mentioned above. In Section 2, we
introduce the papers related to the theme of this
paper. In Section 3, we describe the search algorithm,
system, and prototype. In Section 4, we wrap up the
paper with our conclusions and directions for future
work.
2 RELATED WORK
2.1 Blog Search
As blogs have been spotlighted as a new resource to
search for information (Fujiki, Nanno, Suzuki, and
Okumura, 2004), several search engines perform the
search only for blogs, and there are research papers
only on the topic of blog searching.
First, one paper entitled “A Study of Blog Search”
shows the differences between blogs and general
web pages in detail (Gilad, and Maarten, 2006). It
focuses on queries to find unique features of blogs
that one might apply in a blog search. This paper
surely adds to blog searching studies.
BLOGRANGER offers many kinds of searches
(Fujimura, Toda, Inoue, Hiroshima, Kataoka, and
Sugizaki, 2006). It helps users to reach the area that
they intend to by presenting topics, blogger,
reputation search, and navigational, informational
and transactional searches. Its concept comes from
the view that blogs are personal and thus different
from other web-pages. One paper is on the use of the
concept of a keyword map and feedback (Takama,
Kajinami, and Matsumura, 2005). It consists of a
keyword map as a method of providing feedback.
When a user searches with a keyword, he or she
gives words related to the keyword as feedback. The
user may select the word as the domain that the user
intends to look into, and this gives the user more
precise search results corresponding to the user’s
needs. Finally, there is a paper that discusses using
tags in blog classification (Aixin, Maggy, and Ying,
2007). In this paper, they analyze tags to find
popular tags. They analyze the frequency of these
popular tags to give a weight to each blog. This
paper’s approach is pretty similar to the one we use
in this paper. They use tags created by bloggers to
avoid a general search engine’s categorization, and
they try to classify blogs. However, after analyzing
the tags, they use statistical standards to evaluate
blogs, and such tags cannot include all the various
ranges of blog domains. We attempted to make the
category right for that keyword only.
2.2 K-means
K-means (MacQueen, 1967) is a clustering
algorithm based on distances between instances. It
divides data into K number groups based on a
certain feature of the data. After dividing the data
into K groups, it finds a centre value for each group.
Then, it connects each instance of data to the closest
centre point to make a new group. After dividing the
data into new groups, it repeats the progress of
finding new centre values for each group again and
again. This regeneration is continued until it satisfies
the critical point. Through this process, it tries to
find the most suitable classification form of the
given data.
In this paper, there are three main issues
concerning the use of K-means. One is how to
decide the value of K. Another is how to change
titles to values in order to make it possible to use K-
means. The third is how to decide the end point of
K-means. These three issues are very important in
determining the effectiveness of the use of K-means,
and how we evaluate K-means dealing with these
issues will decide the entire set of features of our
system. Considering these three issues, we
developed our system and algorithm. We introduce
the steps in solving these issues in section 3.
3 SYSTEM & ALGORITHM
3.1 Problem Solving
The final object of every search engine is to provide
the information intended by the user perfectly. That
is why every search engine has its own search
algorithm, weight and search classification. For
example, the search engine ‘Yahoo’ (Yahoo)
provides its own directory search organization, with
categories like image search, job search, product
search, audio search, and so on. The search engine
‘Google’ (Google), like Yahoo, also provides its
own categories. However, since current web pages
have become more specific and have been created in
a wide variety of ways, current general search
engines have an evident limitation with static
directories, especially concerning blog searches.
Because a blog is managed by an individual blogger,
the blogger uploads whatever he or she wants, in
whatever format. So, every keyword in a blog has its
own themes and domains totally different from
keywords of other blogs and other types of web
pages. Users do not use a blog search to shop or to
find a job or a map. If some current engines try to fit
ICEIS 2009 - International Conference on Enterprise Information Systems
62
blog information into current static categories, there
is a possibility of misclassifying and thus failing to
find an appropriate category. In this case, the
categories made for users can limit users from
finding proper data. Therefore, in order to attempt to
satisfy various users’ needs, we analyzed the
contents from the keywords, without any other data.
After analyzing the search results, we let users know
all the major issues related to those keywords by
ranking them in a clustered form. In addition, we
showed users the relative importance of each group.
With this system, we expected users to recognize a
keyword trend faster than ever.
As we mentioned in section 1, problems remain
which current search engines cannot solve. To solve
some problems, we present the search algorithm that
satisfies the following needs.
1. If searched results have the same keyword, can
we cluster this information not with a static category
but with a certain keyword’s distribution?
2. Can we figure out the most important, the most
popular and the most necessary information based
on rank?
3. Can we see the information without overlapping
and grasp the point at first sight?
We do not think that these suggestions will
satisfy all the blog search users’ needs. We also are
not saying that every search engine should follow
our algorithm or change theirs. All we are saying is,
at least, this kind of access can be the one big
improvement which can add power to current search
engines and to the clustering concept discussed in
the search engine section.
3.2 System Architecture
The prototype of the program used to test the
algorithm based on this paper is similar to the
following.
The whole system is managed like the schematic
in figure 1, as you see below, and how each part of
the prototype works is explained from 3.2.1 to 3.2.5
3.2.1 From System User to Search Engine
A user inputs a keyword just like any user does in
existing search engines. Then we send the keyword
to the normal search engine. In this case, we used
the naver blog search engine (Naver), which is the
most famous search engine in Korea. This search
engine performs in the way it normally does.
Figure 1: Architecture of Prototype System.
3.2.2 From Search Engine to Memory
After the blog search is finished, we extract the
search data served by the naver blog search engine
and send it to our prototype’s memory. Through this
process we simply extract the naver blog search
engine’s results without performing any editing.
3.2.3 Memory and Classification System
When the memory provides the search results to the
classification system, it edits the search results to
perform our classification algorithm. It detaches
only titles from the search results to follow our
classification system, which depends on the
similarity between titles, and saves the titles in our
data structure which we describe in the next section.
Through this process, the search results provided by
the general search engine is ready for our system.
When this process is completed, the memory sends
the newly created data structure to the classification
system.
After the classification system clusters the search
results, it sends its results to the memory. The
detailed process of clustering the search results in
the data structure is described in the next section.
The classified results are also returned along with
the data structure.
3.2.4 Memory with Ranking System
The memory sends the classified results provided by
the classification system to the ranking system. The
ranking system consists simply of a method which
ranks the result groups based on their size.
The ranking system returns its ranked result to
the memory when the ranking process is completed.
BLOG CLASSIFICATION USING K-MEANS
63
3.2.5 From Memory to User
After receiving the ranked results from the ranking
system, the memory sends the final results to the
user.
3.3 Definition of Key Algorithm
At first, we would like to present the basic
definitions of our system which are used in this
paper. These definitions indicate a form of data on
the memory from the time of extracting the title
from the search engine to the time the ranking
system ranks the result.
Definition 1. Let T = {
,
,
. . . ,
} be the title
set. When we receive the titles, we name each of
them as
,
,
,
. . . ,
. For instance, if we receive
100 titles, then |T| = 100.
Definition 2. To earn a measurement from a title, we
divide each title into morphemes. So, each title has
its own morpheme group. We define this group as a
keyword set described like KS = {
,
, . . . ,
}. The number n is the same as the n in
Definition 1, since KS is a morpheme set of T. In
addition to this, since each title consists of words,
we define each title KS as KS = {
,
, . . . ,
}. The number m depends on the morpheme
number of each T.
Definition 3. From KS, we create a keyword pool as
KP, described as KP = {
,
, . . . ,
}.
Basically, it collects the keywords from KS to make
the entire keyword set without any keyword
repetition. Therefore, KP has every keyword
included in the search results. The number x means
the number of all the morphemes the search engines
have.
Definition 4. = {
| i = 1, 2, … , n , j =
1,2, … , x } is the keyword vector which tells us that
a certain title has a certain morpheme. Every single
value of
is 0 or 1. For example, when
comparing KS with KP, if
is matched with
we set the
vector value as 1. The number
a could be any number which is in the range from 1
to n. The number b or c could be any number which
is in the range from 1 to x. The number x means the
number of all the morphemes the search engines
have.
3.3.1 Title Analyzing
Since we chose the titles of the analyzed blog as the
classification measure, we detached titles from the
search results as the first step. This was not that hard
to do if we could see the structure of the pages by
using an HTML parser. The other parts of the pages
are linked to the title of each page and are added to
the final result.
3.3.2 Morpheme Analyzing
Before we adjusted the K-means to the search result
served by the general search engine, we had to fit the
form correctly for K-means. K-means requires a
certain heuristic to divide each instance into groups.
We have selected this form as vectors, as you can
see in 3.3.1
To give some values to each title, we decided to
divide the titles into morphemes and give them
values of 1 or 0. We used a morpheme analyzer to
extract morphemes from each title. In this process
we erased each blog’s unique characters, such as an
emoticon.
3.3.3 Keyword Pool Making
In order to make a standard keyword, we created a
keyword pool which had all keywords included in
the search results. This was an essential task in our
algorithm because we had to perform a vector
product with each title in the search results, and this
task made that possible. The process of making a
keyword pool is quite straightforward. Comparing
all the elements in KS, we can simply delete those
that overlapped.
3.3.4 Keyword Vector Making
Finally, we made a keyword vector in order to give a
specific measurement for performing K-means.
Through this progress, we could calculate the
distance between two tiles which reflected the
similarity between them. And this similarity, the
result of the vector product, plays an important role
in classifying titles.
Making a keyword vector is not that complex a
process. Adding to the explanation in 3.3’s
definition 4, here is a more concrete example. Let’s
suppose that there are only five morphemes in KP
which can be described as KP = {a, b, c, d, e} and
there are also two KSs which can be described as
= {a, b, e} and
= {a, c, d}. Then, the KV
of
and
will be {1, 1, 0, 0, 1} and {1, 0, 1,
1, 0}.
ICEIS 2009 - International Conference on Enterprise Information Systems
64
3.4 K-means
3.4.1 First Classifying and Deciding K Value
When using K-means as a classification method,
deciding the value of K is one of the most important
issues. Although most cases using K-means set K as
a static value, we set K as a dynamic value because
we wanted to reflect the fact that each keyword has
absolutely different kinds of domains and quantities.
It makes every searched word have its own K. The
process of deciding the K value is easy to understand.
During the process of classification, we checked
every KS in detail. The first one,
, should consist
of the first result group since there are no prior result
groups. Following that, one performs vector product
with the existing result group. Then, the following
KS is added to the result group, thus making the
biggest vector product value with it. If the vector
product values performed with all the existing result
groups are all zero, the following KS creates its own
result group since the following KS has no similarity
with the others. Finally, the number of result groups
is the number of this searched word’s K.
3.4.2 Centre Value and Classification
K-means is based on a reputation, so we should
decide a centre value for each result group and
reclassify all KSs into the new result groups. The
centre value of each group is easy to find since we
have each KSs vector value KV. Just add up all the
vectors in each result group. Then divide each sum
by the number of each group’s KVs. This gives a
perfect centre value for each result group. So, we
added all KVs to the new group, which made the
biggest vector product value with this group’s centre
value. After completing the classification process,
we simply repeated this process from the beginning.
3.4.3 Deciding the End Point
Deciding the critical value to end K-means is also an
important issue concerning K-means. Through
3.4.2’s process, each search result may find its
perfect position. Therefore, when there were no
more changes in every result group, although K-
means did its reputational classifying work, we
stopped doing K-means. This offered the best result
groups to the users.
3.5 Ranking System
Presenting search results with rank and distribution
is also an important aspect of our system. The reason
that we provide rank with distribution is that with
rank only users cannot figure out the importance of
each group completely. Among the classified groups,
we calculate each group’s size first. The largest one
is the first one to be presented since the largest
group can be considered as the most spotlighted one.
The distribution is served in the form of a percentage.
The percentage provides an additional relative
importance of that group. Therefore, when users
receive the final results of our system, users can
understand the degree to which the subject and how
many subjects are the most important in this
keyword’s area.
4 EVALUATION
In order to evaluate section 3’s algorithm and system,
we made a prototype which can perform real–time
blog searches and estimate the effectiveness of the
prototype. The prototype system’s range is on
Korean blogs and the estimation method is based on
the CSIM method (Chung and Lee, 2001).
Figure 2: Translated Screen Image of Prototype System.
The prototype system demonstrates its result, as
shown in Fig. 2. We performed this test in Korean
but we translated Fig 2 into English to help you to
see how the process works. First, it shows its result
only by the group’s title which can represent that
group, and when a user clicks on the button
following the title, a new panel shows the user the
entire data set of the group. If the user clicks on a
certain title in the result panel, the system links the
user to that page. With such an interface a user can
figure out the distribution of the data immediately.
Cluster SIMilarity (CSIM) is derived from
Rand’s method (Rand, 1971) which is often used in
estimating clustered group similarity. It is a method
BLOG CLASSIFICATION USING K-MEANS
65
which applies Dice’s coefficient to Rand’s method
in order to overcome some defects of Rand’s method.
For the three types of web pages (Border, 2002),
we decided to choose queries from informational
classes which are estimated as the most proper ones
for blog searches (Gliad and Maarten, 2006). We
chose one query each from movie, music, and book
categories. First, ‘X-Japan,’ the name of one of the
most famous rock bands in Japan, is the query from
music. Second, we chose ‘Cha T.H,’ who is one of
the most famous actors in Korea, as the query word
from movies. Lastly, ‘Ekuni Gaori,’ who is one of
the most famous Japanese novelists, is the query
from books.
We tested 50 sets of blog data from the Naver
search engine, without performing any editing.
Although our prototype system could test its
algorithm with all blog pages on the web, we
decided to test just 50 pages this time. We will test
more pages for more accuracy later. Before our
prototype system perofrmed its task, we made an
ideal clustered set by hand. After the prototype
system created its result set, we compared this with
the ideal set by CSIM.
Table 1: The evaluation result.
Query word X-Japan Cha T.H. Ekuni Gaori
CSLM value 0.857 0.711 0.805
Since the CSIM value is quite close to 1, we can
conclude that the prototype system is successful in
clustering blog information. Although the test was
performed on only 50 sets of blog data, it certainly
clustered data which should be clustered, so we
think that the larger the example set is, the more
exact the results will be. We expect that this system
will be able to offer useful and special information
to users and companies that want to know the
public’s response to their products or image.
5 CONCLUSIONS
In this paper, we discuss a blog search algorithm that
considers the characteristics of blog content based
on the assumption that the resultant blog
classification can provide more valuable information
to users. We also made a simple prototype to
evaluate our algorithm. In order to test this system,
we tried to find features of a blog and the problems
of general search engines, and then find a solution
which could solve those problems to an extent. We
decided to use the concept of K-means as the
classification method. We developed our own
algorithm to adjust K-means to blog information. As
shown in section 4, our algorithm and system
provides certain benefits to users with clustered
groups. It may not satisfy all the users, but it can
give additional useful data to users and suggest a
new approach to the blog search engine field.
For future research, there is something else to
consider. There were three important issues in
making an algorithm with K-means, as you can see
in section 2.2, and we do not think that our solution
suggested in this paper is the only possible one. So
we will try to find the best solution which can
extract a better weight from the blog and choose a
better K and critical point. In addition, we will study
more classification methods which can be matched
more closely with blog searches. Finally, nowadays
a variety of search algorithms and methods used in
search engines exist. Since our final goal is to
present the best blog algorithm, we will study other
search mechanisms, including classification.
ACKNOWLEDGEMENTS
This research was financially supported by the
Ministry of Knowledge Economy(MKE) and Korea
Industrial Technology (KOTEF) through the Human
Resource Training Project for Strategic Technology.
REFERENCES
Aixin, S., Maggy, S., Ying, L. 2007. Blog Classification
Using Tags: An Empirical Study. In ICADL 2007.
Bloglines: http://www.bloglines.com/.
Blogpulse: http://www.blogpulse.com/.
BLOGRANGER: http://ranger.labs.goo.ne.jp/.
BlogWatcher: http://blogwatcher.pi.titech.ac.jp/.
Broder A. 2002. A Taxonomy of Web Search. In SIGIR
Forum.
Chung, Y.M., Lee, J.Y. 2001. A corpus-based approach to
comparative evaluation of statistical term association
measures. In J. of the American Society for
Information Science and Technology.
Fujiki, T., Nanno, T., Suzuki, Y., Okumura, M. 2004.
Identification of Bursts in a Document Stream. In First
International Workshop on Knowledge Discovery
2004.
Fujimura, K.,Toda, H., Inoue, T., Hiroshima, N., Kataoka,
R., Sugizaki M. 2006. BLOGRANGER – A multi-
faceted Blog Search Engine. In WWW 2006.
Gilad, M., Maarten, R. 2006. A Study of Blog Search. In
ECIR 2006. LNCS 3936.
Google, http://www.google.com/.
Kumar, R., Novak, J., Raghavan, P., Tomkins, A. 2003.
On the bursty evolution of blogspace. In WWW’03:
ICEIS 2009 - International Conference on Enterprise Information Systems
66
Proceedings of the 12
th
international conference on
world wide web. ACM Press.
Macqueen J. 1967. Proceedings of the Fifth Berkeley
Symposium on Mathematical Statistics and Probability,
Volume 1: Statistics. University of California Press.
Mukul, J., Nikhil, B. 2006. BlogHarvest: Blog Mining and
Search Framework. In COMAD 2006.
Naver: http://www.naver.com
Rand, W.M. 1971. Objective Criteria for The Evaluation
of clustering Methods. In J. of the American Statistical
Association.
Takama, Y., Kajinami, T., Matsumura, A. Application of
Keyword Map-based Relevance Feedback to
Interactive Blog Search. In IEEE 2005.
Technorati Weblog: State of the Blogsphere,
http://technorati.com/weblog/2006/02/83.html
Yahoo, http://www.yahoo.com/.
BLOG CLASSIFICATION USING K-MEANS
67
