Web Page Classification Based on Web Page Size and
Hyperlinksand Website Hyperlink Structure
Denis L. Nkweteyim
School of Information Sciences, 135 N. Bellefield Avenue, University of Pittsburgh,
Pittsburgh, PA 15260
Abstract: This paper presents a new metric, Page Rank × Inverse Links-to-
word count Ratio (PR × ILW), used in classifying web pages as content or
navigation. The metric combines web page size and number of hyperlinks on a
page, and the page rank metric based on website topology, to compute the new
metric. We present a theoretical basis for the new metric, and the results of a
web page classification study, which show that the new metric, when combined
with the links-to-word count ratio of web pages, accurately classifies the pages
into the two categories.
1 Introduction
Web page classification is required in several situations. For example, in web content
mining and web usage mining [1, 2], web pages are grouped based on their content
(web content mining), or classified as content or navigation pages (web usage
mining). In hypermedia personalization systems, it is often necessary for the system
to implicitly deduce items of interest to the user without explicitly requiring the user
to rate items, that way reducing the cognitive load placed on the user. Automatic
classification of web pages is useful in such systems.
There has been much research on parameters that can be used to automatically
deduce items of interest to users in personalization systems. Reading time, for
example, has been extensively studied as a factor that determines interesting pages in
hypermedia (see for example, [3-7]). Chen, Park, and Yu [8] proposed Maximal
Forward Reference (MFR) as an implicit indicator of user interests in web pages as
they navigate a web site. This approach is based on tracking users’ forward
references (pages not in the set of pages already seen) and backward references (pages
already in the set of visited pages), as they navigate a web site. Other implicit interests
indicators that have been researched include: edit wear and read wear [9], and
examination, retention and reference [10, 11].
Our research involves the design of hyperlink recommender systems based on past
user navigation behaviors. The work involves mining of frequent user access patterns
from web user access logs, and making recommendations based on these patterns.
We use a number of heuristics to minimize errors as individual user sessions are
extracted from the web logs (see [2, 8, 12-14] for a discussion of heuristics used), and
the MFR heuristic [8] to extract transactions within the session data. We aim to
L. Nkweteyim D. (2005).
Web Page Classification Based on Web Page Size and Hyperlinksand Website Hyperlink Structure.
In Proceedings of the 5th International Workshop on Pattern Recognition in Information Systems, pages 225-233
DOI: 10.5220/0002577102250233
Copyright
c
SciTePress
discover transactions of lengths between 2 and 5 URLs and mine these transactions
for association rules that show the correlation relationships between user navigation
behaviors and the pages they find interesting. The choice of desirable transaction
lengths was motivated by research [15], that suggests that most users follow between
two and five hyperlinks before reaching a page of interest.
Using MFR alone however, we found out that there were several transactions much
longer than 5 URLs. We assumed that these long transactions contained hidden
content (or recommendable) pages, which the MFR heuristic was not able to find.
The challenge therefore, was to find a heuristic that could be used to make plausible
guesses on what these hidden content pages were. Our approach combined the ratio
links count:term count of web pages and the topology of the web site to compute a
number that determined whether a page within a very long transaction should be
classified as content or navigation, and hence determine new transaction boundaries.
2 Parameters Used to Determine Content and Navigation Pages
Based on Web Page Characteristics. A content page is a page with information that
could be of interest to a user as she browses a web site. In general, web page contents
could be text, graphic, or multimedia. For the purpose of this research though, only
text was treated as content because text can be much more easily manipulated. The
size of a web page can be considered equal to the number of terms found on the page.
The larger the page size, the more likely it is that the page is a content page. In a
hyperlink recommender system, the objective may be to recommend interesting
content-rich pages to the user. A web page also typically has one or more hyperlinks.
In general, the larger the number of hyperlinks on a web page, the more likely it is
that the page is a navigation page.
Because a web page usually comprises text and hyperlinks, most web pages exhibit
both navigation and content properties, and one or more metrics other than page size
and the number of hyperlinks present, are required to correctly classify them.
Besides, other factors may influence the classification of web pages (for example, the
links to word count ratio, LW). Intuitively, the larger the value of LW, the more
likely it is that the page is a navigation page, and vice versa for content pages. It
should be noted though, that some web pages with high LW values could be content
rich; likewise, some pages with low LW values may have too little content.
Based on Web Site Topology. In a navigation task a user follows hyperlinks the
labeling of which suggest to the user that she is moving to pages that will meet her
current information needs. Hyperlink labels are conceived by the web site designer to
serve as local cues that users process in making judgments on which hyperlinks to
follow. In foraging theory [16-19], these cues are called information scent
1
, and could
serve as an additional page classification parameter. In this research, we measured
information scent using a variation of Google’s PageRank algorithm.
1
Information scent characterizes how users—like organisms that use scent to determine where
go next—evaluate the utility of their actions to lead them to target information. The greater
the scent, the more likely is the user to access the information.
226
PageRank PR, [20, 21] is a number that Google uses to determine the importance
of a web page, and is one of several parameters used to determine the ranking of
pages returned by the Google search engine. The PR algorithm assumes that a
hyperlink from one page to another is a vote from the former to the latter. The
importance of a page is determined by the number of votes it receives, and the relative
importance of the pages casting the votes. A page also loses some of its page rank
based on the number of outgoing links on the page. The following equation is used to
determine the page rank of Page A:
))(/)(...)(/)(()1()(
11 nn
TCTPRTCTPRddAPR +++−=
(1)
where, PR(T
I
) is the rank of page I; C(TI) is the number of hyperlinks on page I;
PR(TI)/C(TI) is the contribution of page I to the rank of page A, if there is a link from
page I to page A; and d is a damping factor, usually set to 0.85. In this work, PR was
computed through an iterative process, with the page ranks on the right hand side of
the equation initialized to 1/N, where N is the number of web pages in the web site,
and d set to 0.85.
In a navigation task, PR can be considered to be reflective of the web site
designer’s view of the way the site should be navigated: web pages with many
incoming links are deemed more important than web pages with few incoming links.
The problem with page rank in the domain of web navigation is that it is based solely
on the network of links between web pages; it does not predict correctly how real
users navigate the web site, and it says nothing about the content or size of pages.
Hence, a page with several links pointing to it is likely going to be judged important,
even if its content is minimal (and hence, not recommendable), while a page with
several outgoing links is likely to lose much of its page rank, even if its content is
important.
Combining PR and LW. Working with the intuition that a web page with a high page
rank, but with little content is not as worth recommending, as a page with a low page
rank, but with much content, we sought to combine PR and LW to get a single metric
to use to classify web pages.
Let ILW represent the inverse links-to-word count ratio (i.e., 1/LW). Table 1
suggests that PR × ILW is a better metric to use to determine the content rating (how
likely it is that a page is a content page) of a page than PR alone. For example, the
third row of the table represents pages with low LW values (i.e., high ILW values,
and hence high content rating), and large incoming:outgoing links ratio (i.e., high
PR). The corresponding PR × ILW is equivalent to a large increase in the PR score
since both PR and ILW are large. On the other hand, for the fourth row, which
corresponds to web pages with high PR and low ILW, there is only a small increase in
the PR × ILW score when compared to PR. A similar argument can be made for all
the other rows of the table.
3 Experiment
The objective of this experiment was to determine how useful the LW and PR × ILW
scores of web pages can be in the classification of the pages as content or navigation.
227
Subjects. The subjects comprised three Information Sciences graduate students
who were very familiar with the web site of the School of Information Sciences,
University of Pittsburgh (SIS): two fifth year and one third year.
Method. The web site of SIS (http://www.sis.pitt.edu) was used to test the new
metric. A crawler was used to search for the hyperlinks found on the site’s HTML
pages, and the corresponding pages parsed to remove tokens related to page
presentation, leaving only the contents. For each page, the SMART stop list [22] was
used to remove commonly used words and the Porter stemming algorithm [23] to
stem each term. High and low frequency terms (occurrence frequency respectively
greater than 25%, and less than 1% of the web pages) were also removed. The web
site topology was then constructed using an adjacency-lists graph representation.
Table 1. Illustrating the moderating effect of ILW on PR in the PR × ILW metric
Outgoing
Links
Incoming
LW
Content
PageRank
PR
×
ILW
L
L
L
H
M
M
++
L
L
H
L
M
M
+
L
H
L
H
H
H
++
L
H
H
L
H
H
+
H
L
L
H
L
L
++
H
L
H
L
L
L
+
H
H
L
H
M
M
++
H
H
H
L
M
M
+
L = Low ; H = High; M = Moderate; + = small increase; ++ = large increase
Fig. 1. Screenshot of one of the web pages that subjects classified. The top panel contains
directions to subjects, the middle panel the page currently being classified, and the bottom
panel page URLs and their ratings
228
Next, the page contents and site topology were used to determine PR and PR ×
ILW for each HTML page. The URLs of 4 web pages were then randomly assigned
without replacement to each of the following 9 groups: PR
high
PR
med
, and PR
low
(PR of
the page is among the top, middle, and low third respectively of page ranks); ILW
high
,
ILW
med
, ILW
low
, for the ILW metric; and PR × ILW
high
, PR × ILW
med
, PR × ILW
low
,
for the PR × ILW metric. Finally, all 36 URLs were presented as hyperlinks on a web
page to the subjects, who were instructed to view and rate the corresponding web
pages on a scale of 0 to 10 using three different scales: a content scale, a navigation
scale, and a dual content/navigation scale. Subjects' responses were collected in a
web-based form containing their ratings data. Figure 1 shows a screenshot of one of
the rating screens presented them.
4 Results and Discussion
Figure 2 shows the page ratings on the three scales by the subjects (S1, S2, and S3).
As can be seen, the ratings were largely consistent, except for some ratings by S2,
which we believed were errors (see the comments in the last column of the figure).
For these erroneous cases, the relevant ratings were not used in computing mean
ratings.
Figure 3 shows the variation of average user ratings with LW. As expected, web
pages were generally classified as content (high content and dual scale ratings) for
low values of LW (left of graph), and as navigation (high navigation scale ratings) for
high LW (right of graph). Also as expected, navigation ratings were strongly
negatively correlated with content ratings, while the content and dual scale ratings
were strongly positively correlated, and so a single scale would be sufficient. Finally,
Figure 3 suggests that LW alone is not sufficient to separate content from navigation
pages. For example, there are a number of pages with relatively high LW ratio that
subjects classified as content pages, but that using the LW parameter alone, would be
classified as navigation pages.
Table 2 shows subjects’ mean content scale ratings, and the page LW and PR ×
ILW values arranged first in order of LW (Columns 1–4 ), and then in order of PR ×
ILW (Columns 5–8 ). From the table, it can be seen that in general, web pages with a
LW value below 0.05 or a PR × ILW value above 1.8 can be classified as content
pages. We notice that using LW alone, content pages 35, 8, 7, 1, 2, and 4 would
normally be classified as navigation pages because their LW values are above the
minimum threshold of 0.05. However, because the corresponding PR × ILW values
are above the 1.8 threshold for classifying pages as content pages, these pages are
correctly classified.
229
Content Scale
Navigation Scale
Dual Scale Comments
Item No
S1 S2 S3 S1 S2 S3 S1 S2 S3
1
5
5
5
5
5
3
5
10
6
d
S2: dual scale rating inconsistent
2
8
3
4
2
7
4
7
2
5
3
10
10
10
0
0
0
10
10
10
4
9
7
7
1
2
7
9
6
5
6
2
10
2
8
0
8
2
10
2
a
S2: Must have missed long list of hyperlinks
at bottom of page
7
9
8
7
2
1
2
8
8
7
8
8
8
4
3
1
6
7
7
4
9
1
8
1
6
1
10
3
8
1
a
S2: Page comprises almost entirely of
hyperlinks
10
9
2
9
7
2
8
8
2
9
c
S2: Content rating too low; page comprises
mainly plain text
11
1
1
1
10
9
10
0
1
0
12
10
7
9
1
1
1
9
7
10
13
10
8
9
3
1
3
8
8
9
14
8
8
7
7
1
4
5
8
7
15
10
10
10
0
0
0
10
10
10
16
9
10
10
2
0
1
9
10
9
17
9
10
10
1
0
1
9
10
10
18
5
10
6
8
0
4
4
10
5
a
S2: All ratings inconsistent; large table with
many links and moderate amount of text
19
9
9
6
2
1
3
9
9
7
20
9
9
8
2
1
2
9
9
8
21
8
10
7
3
0
7
8
10
5
22
7
10
9
2
0
2
8
10
9
23
10
9
5
0
0
1
9
9
9
24
10
9
9
0
0
1
10
9
10
25
1
10
1
9
0
9
1
10
0
a
S2: All ratings inconsistent; large table with
many links and moderate amount of text
26
1
10
1
9
0
10
1
10
0
a
S2: All ratings inconsistent; large table with
many links and moderate amount of text
27
9
10
8
1
0
2
9
10
9
28
10
5
7
2
5
7
9
5
5
30
9
5
2
2
5
8
8
5
1
31
10
10
10
0
0
0
10
10
10
32
7
6
7
3
4
3
6
6
7
33
5
10
2
9
0
3
2
10
3
34
9
8
10
1
1
2
9
8
10
35
9
9
10
1
1
3
9
9
9
a – all 3 of subject’s ratings not used in computing average score; d – su
bject’s dual scale rating not used in
computing average score; c - subject’s content scale rating not used in computing average score
Fig. 2. Subject ratings of the web pages presented in the experiment. The commented items
refer to cases where there were large discrepancies in subject ratings. Results of 3 of the 36
web pages that were presented to users are omitted because these pages were no longer linked
to by the time the results were analyzed
0
2
4
6
8
10
12
0
0
0
.01
0
.01
0.01
0.0
3
0.0
3
0.03
0.04
0.07
0.07
0.08
0.1
0
.11
0
.14
0
.17
0
.4
Outgoing Hyperlinks:Word Count ratio (LW)
Mean Page Rating
Content Scale Navigation Scale Dual Content/Navigation Scale
Fig. 3. Variation of page ratings with LW on the content, navigation and dual
content/navigation ratings scales
230
Table 2. Combining LW and PR × ILW thresholds to classify web pages as content or
navigation
Sorted by LW Sorted by PR × ILW
PageID
Mean Content
Rating
LW PR × ILW PageID
Mean Content
Rating
LW PR × ILW
31 10.00 0.00 6.20 2 5.00 0.14 7.30
15 10.00 0.00 6.14 31 10.00 0.00 6.20
3 10.00 0.00 2.63 15 10.00 0.00 6.14
28 7.33 0.01 3.80 12 8.67 0.01 5.95
12 8.67 0.01 5.95 23 8.00 0.01 5.84
22 8.67 0.01 3.77
8 6.67 0.07 5.55
17 9.67 0.01 3.86 24 9.33 0.01 5.05
23 8.00 0.01 5.84 13 9.00 0.02 4.68
24 9.33 0.01 5.05 34 9.00 0.04 3.88
13 9.00 0.02 4.68 17 9.67 0.01 3.86
14 7.67 0.03 2.65 28 7.33 0.01 3.80
21 8.33 0.03 2.57 22 8.67 0.01 3.77
16 9.67 0.03 2.61
1 5.00 0.09 3.48
32 6.67 0.03 2.15 35 9.33 0.07 3.44
27 9.00 0.03 2.58 10 9.00 0.05 3.37
20 8.67 0.04 1.85
4 7.67 0.17 3.27
34 9.00 0.04 3.88 7 8.00 0.08 3.12
10 9.00 0.05 3.37 6 2.00 0.07 2.80
35 9.33 0.07 3.44 14 7.67 0.03 2.65
8 6.67 0.07 5.55 3 10.00 0.00 2.63
6 2.00 0.07 2.80 16 9.67 0.03 2.61
*18 5.50 0.08 1.36 27 9.00 0.03 2.58
7 8.00 0.08 3.12 21 8.33 0.03 2.57
1 5.00 0.09 3.48 32 6.67 0.03 2.15
26 1.00 0.10 1.28 20 8.67 0.04 1.85
*30 5.33 0.11 1.62 *30 5.33 0.11 1.62
25 1.00 0.11 1.21 11 1.00 0.34 1.51
2 5.00 0.14 7.30 *18 5.50 0.08 1.36
*33 5.67 0.14 1.08 26 1.00 0.10 1.28
*19 8.00 0.16 0.97 9 1.00 0.40 1.23
4 7.67 0.17 3.27 25 1.00 0.11 1.21
11 1.00 0.34 1.51 *33 5.67 0.14 1.08
9 1.00 0.40 1.23 *19 8.00 0.16 0.97
Content pages marked with an asterisk, “*” (18, 30, 33, and 19) were not correctly
classified. We consider each of them in turn. Pages 18, 30 and 33 are borderline
content pages (ratings close to 5), and they happen to have only 1, 2 and 1 incoming
links respectively. Because these pages are linked to from so few pages, users have
very few opportunities to reach them, and so not treating them as content or
potentially recommendable pages is desirable.
Page 19 on the other hand was rated very highly as a content page, but both LW
and PR × ILW failed to classify it as such. The reason for is that its PR, and thus
PR × ILW, is too low to change its classification to content. The low PR resulted
from the fact that the page is linked to from only one other page in the web site.
Again, not classifying this page as a content page for recommendation purposes is
desirable because users have very little opportunity to ever get to the page when they
browse the web site.
Finally, navigation Page 6 which was correctly classified using LW is now
incorrectly classified using PR × ILW.
231
5 Conclusion
We have shown that PR × ILW is useful in classifying web pages as content or
navigation in applications where content pages are pages that a user browsing a web
site may find interesting. This information can be used by a browsing agent that helps
the user by observing her navigation behavior, comparing that behavior with those of
past users of the web site, and recommending to her the content pages that the past
users found interesting.
The PR × ILW metric works well in this regard because user browsing behaviors
are usually constrained by the link structure of a web site (users typically navigate a
site by following hyperlinks on pages on the site), and the metric exploits both this
link structure and the properties of individual web pages.
It is worth noting that this classification scheme may not be as useful in other
domains, for example where the main interest is in the contents, and not necessarily
the connectedness of web pages.
References
1. Mobasher, B., R. Cooley, and J. Srivastava: Automatic Personalization Based on Web
Usage Mining. In Communications of the ACM (2000) 142–151
2. Cooley, R., B. Mobasher, and J. Srivastava: Web Mining: Information And Pattern
Discovery on the World Wide Web. In International Conference on Tools With Artificial
Intelligence, Newport Beach, CA, (1997) 558–567
3. Morita, M. and Y. Shinoda: Information Filtering Based on User Behavior Analysis and
Best Match Text Retrieval. In Seventeenth annual international ACM-SIGIR conference
on Research and development in information retrieval (1994) 272–281
4. Konstan, A..J., Miller, N., Maltz, D., Herlocker, L., Lee, R., Riedl, J.: GroupLens:
Applying Collaborative Filtering to Usenet News. In Communications of the ACM, ACM
Press, New York (1997) 77–87
5. Kim, J., W.D. Oard, and K. Romanik: User Modeling for Information Filtering Based on
Implicit Feedback. In ISKO-France 2001, Nanterre, France (2001)
6. Claypool, M., Le, P., Waseda, M., Brown, D.: Implicit Interest Indicators. In Proceedings
of the 6th International Conference on Intelligent User Interfaces, Santa Fe, NM (2001)
33–40
7. Kelly, D. and Belkin, N.: Reading Time, Scrolling and Interaction. In Proceedings of the
24th Annual International ACM SIGIR conference on Research and Development in
Information Retrieval (2001) 408–409
8. Chen, M.S., Park, J.S., Yu, P.S.: Data Mining for Path Traversal Patterns in a Web
Environment., in Proceedings of the 16th International Conference on Distributed
Computing Systems (1996) 385–392
9. Hill, C.W., Hollan, D. J., Wroblewski, D., McCandless, T.: Edit Wear and Read Wear. In
Proceedings of Conference on Human Factors and Computing Systems Monterey, CA
(1992)
10. Nichols, D.: Implicit Rating and Filtering. In Proceedings of the Fifth DELOS Workshop
on Filtering and Collaborative Filtering Budapest, Hungary (1998) 31–36
11. Oard, W.D., Kim, J.: Implicit Feedback for Recommender Systems. In AAAI Workshop
on Recommender Systems Madison, WI (1998)
12. Pitkow, J.: In Search of Reliable Usage Data on the WWW. In Proceedings of the Sixth
International WWW Conference (1997) 1343–1355
232
13. Pirolli, P., Pitkow, J., Rao, R.: Silk from a Sow's Ear: Extracting usable structures from
the Web. In Conference on Human Factors in Computing Systems (CHI-96) (1996)
14. Cooley, R., Mobasher, B., Srivastava, J.: Data preparation for Mining World Wide Web
Browsing Patterns. Journal of Knowledge and Information Systems 1(1) (1999) 1999
15. Yan, T., Jacobsen, M., Garcia-Molina, H., Dayal, U.: From User Access Patterns to
Dynamic Hypertext Linking. In 5th World Wide Web Conference. Paris (1996)
16. Larson, K. Czerwinski, M.: Web Page Design: Implications of Memory, Structure, and
Scent for Information Retrieval. In CHI'98 Human Factors in Computing Systems. ACM
Press (1998)
17. Pirolli, P.: Computational Models of Information Scent-Following in a Very Large
Browsable Text Collection. In CHI ’97 Human Factors in Computing Systems Atlanta,
GA. ACM Press (1997)
18. Pirolli, P., Fu, W.: SNIF-ACT: A Model of Information Foraging on the World Wide
Web. In: Brusilovsky, P., Corbert, A., De Rosis, F. (eds.): User Modeling 2003 (2003)
19. Card, S.K., Pirolli, P., Van Der Wege, M., Morrison, J. B., Reeder, R.W., Schraedley,
P.K., Boshart, J.: Information Scent as a Driver of Web Behavior Graphs: Results of a
Protocol Analysis Method For Web Usability, in Proceedings of the Conference on
Human factors in Computing Systems, CHI '01. ACM Press, Seattle, WA (2001) 498–505
20. Rogers, I.: The Google PageRank Algorithm and How it Works. Retrieved September 5
2003 from http://www.iprcom.com/ papers/pagerank/
21. Craven, P.: Google's PageRank Explained and How to Make The Most of it. Retrieved
September 5 2003 from http://www. webworkshop.net/ pagerank.html
22. Rocchio, J.: Relevance Feedback in Information Retrieval. In: Salton, G. (ed): The Smart
Retrieval System: Experiments in Automatic Document Processing. Prentice-Hall,
Englewood, NJ.(1971) p. 313–323
23. Porter, M.: An Algorithm for Suffix Stripping. Program. 14 (1980) 130–137
233
