N e u r a l N e t w o r k s

Neural Networks for

Web Content Filtering
Pui Y. Lee and Siu C. Hui, Nanyang Technological University
Alvis Cheuk M. Fong, Massey University, Albany

W ith the proliferation of harmful Internet content such as pornography, violence,

and hate messages, effective content-filtering systems are essential. Many

Web-filtering systems are commercially available, and potential users can download trial

versions from the Internet. However, the techniques these systems use are insufficiently

accurate and do not adapt well to the ever-changing
Web. (For more on current approaches and systems,
The Intelligent see the “Web Content-Filtering Approaches” and
“Current Web Content-Filtering Systems” sidebars.)
Classification Engine To solve this problem, we propose using artificial
neural networks (ANNs)1,2 to classify Web pages
uses neural networks’ during content filtering. We focus on blocking
pornography because it is among the most prolific
learning capabilities to and harmful Web content. According to CyberAtlas,
pornography-related terms such as “sex” and “porn”
provide fast, accurate are among the top 20 search terms queried at the 10
leading Internet portals and search engines.3 Fur-
differentiation thermore, research suggests that pornography is
addictive and causes harmful side effects.4 However,
between pornographic our general framework is adaptable for filtering other
objectionable Web material.
and nonpornographic
Know the enemy
Web pages. The basic How do pornographic Web pages differ from oth-
ers? We attempt to answer this question by studying
framework can also these pages’ characteristics and analyzing data.
serve to distinguish Characteristics
Understanding pornographic Web pages’ charac-
other types of Web teristics can help us develop effective content analysis
techniques. Although it is well known that porno-
content. graphic Web sites contain many sexually oriented
images, text and other information can also help us dis-
tinguish these sites. We focus on three characteristics:
page layout format, use of PICS (Platform for Internet
Content Selection) ratings, and indicative terms.
Page layout format. We treat all the HTML docu-
ments used to construct a multiframe Web page as a
single entity. This is because statistics obtained from
any aspects of a Web page should be derived as a
whole from the aggregated data collected from every
HTML document that is part of that page.
PICS use. Web publishers can use PICS labels to
limit access to Web content (see the “Web Content-
Filtering Approaches” sidebar). A PICS label is usu-
ally distributed with the associated Web page by one
of these methods:
• The Web publisher (or Web content owner) embeds
the label in the HTML code’s header section.
• The Web server inserts the label in the HTTP
packet’s header section before sending the page
to a requesting client.
In both cases, we can’t determine whether a Web
page has a PICS label simply by inspecting the Web
page contents displayed in a browser. The first
method requires viewing the Web page’s HTML
code, which you can easily do with any major Web
browser that supports such functionality. In contrast,
the second method requires checking the HTTP
packet’s header section, which you can’t do with any
known browser. Consequently, to analyze porno-
graphic Web sites’ use of PICS systems, we collect
statistics from the samples’ HTML code.
Indicative terms. Terms (words or phrases) that indi-
cate pornographic Web pages fall into two major
groups according to their meanings and use. Most

48 1094-7167/02/$17.00 © 2002 IEEE IEEE INTELLIGENT SYSTEMS

 Web Content-Filtering Approaches
The four major content-filtering approaches are Platform
for Internet Content Selection, URL blocking, keyword fil-
tering, and intelligent content analysis.
Platform for Internet Content Selection
PICS is a set of specifications for content-rating systems (for
URLs for these rating systems, see the “Useful URLs” sidebar).
It lets Web publishers associate labels or metadata with Web
pages to limit certain Web content to target audiences. The
two most popular PICS content-rating systems are RSACi and
SafeSurf. Created by the Recreational Software Advisory Coun-
cil, RSACi uses four categories—harsh language, nudity, sex, and
violence. For each category, it assigns a number indicating the
degree of potentially offensive content, ranging from 0 (none)
to 4. SafeSurf is a much more detailed content-rating system.
Besides identifying a Web site’s appropriateness for specific age
groups, it uses 11 categories to describe Web content’s potential
offensiveness. Each category has nine levels, from 1 (none) to 9.
Currently, Microsoft Internet Explorer, Netscape Navigator,
and several Web-filtering systems offer PICS support and can
filter Web pages according to the embedded PICS rating labels.
However, PICS is a voluntary self-labeling system, and each Web
content publisher is totally responsible for rating the content.
Consequently, Web-filtering systems should use PICS only as a
supplementary filtering approach.
URL blocking
This technique restricts or allows access by comparing the
requested Web page’s URL (and equivalent IP address) with
URLs in a stored list. Two types of lists can be maintained. A
black list contains URLs of objectionable Web sites to block; a
white list contains URLs of permissible Web sites. Most Web-
filtering systems that employ URL blocking use black lists.
This approach’s chief advantages are speed and efficiency. A
system can make a filtering decision by matching the requested
Web page’s URL with one in the list even before a network
connection to the remote Web server is made. However, this
approach requires implementing a URL list, and it can identify
only the sites on the list. Also, unless the list is updated con-
stantly, the system’s accuracy will decrease over time owing to
the explosive growth of new Web sites. Most Web-filtering sys-
tems that use URL blocking employ a large team of human
reviewers to actively search for objectionable Web sites to add
to the black list. They then make this list available for down-
loading as an update to the list’s local copy. This is both time
consuming and resource intensive.
Despite this approach’s drawback, its fast and efficient oper-
ation is desirable in a Web-filtering system. Using sophisticated
content analysis techniques during classification, the system
can first identify the nature of a Web page’s content. If the sys-
tem determines that the content is objectionable, it can add
the page’s URL to the black list. Later, if a user tries to access
the Web page, the system can immediately make a filtering
decision by matching the URL. Dynamically updating the black
list achieves speed and efficiency, and accuracy is maintained
provided that content analysis is accurate.
Keyword filtering
This intuitively simple approach blocks access to Web sites on
the basis of the occurrence of offensive words and phrases on
those sites. It compares every word or phrase on a retrieved
Web page against those in a keyword dictionary of prohibited
words and phrases. Blocking occurs if the number of matches
reaches a predefined threshold.
This fast content analysis method can quickly determine if a
Web page contains potentially harmful material. However, it is
well known for overblocking—that is, blocking many Web
sites that do not contain objectionable content. Because it fil-
ters content by matching keywords (or phrases) such as “sex”
and “breast,” it could accidentally block Web sites about sex-
ual harassment or breast cancer, or even the home page of
someone named Sexton. Although the dictionary of objection-
able words and phrases does not require frequent updates,
the high overblocking rate greatly jeopardizes a Web-filtering
system’s capability and is often unacceptable. However, a Web-
filtering system can use this approach to decide whether to
further process a Web page using a more precise content
analysis method, which usually requires more processing time.
Intelligent content analysis
A Web-filtering system can use intelligent content analysis to
automatically classify Web content. One interesting method for
implementing this capability is artificial neural networks, which
can learn and adapt according to training cases fed to them.1
(Our Intelligent Classification Engine uses this method—see the
main article.) Such a learning and adaptation process can give
semantic meaning to context-dependent words, such as “sex,”
which can occur frequently in both pornographic and non-
pornographic Web pages. To achieve high classification accuracy,
ANN training should involve a sufficiently large number of train-
ing exemplars, including both positive and negative cases.
ANN inputs can be characterized from the Web pages, such as
the occurrence of keywords and key phrases, and hyperlinks to
other similar Web sites.
Reference
1. G. Salton, Automatic Text Processing, Addison-Wesley, Boston, 1989.

are sexually explicit terms; the rest consist
primarily of legal terms used to establish the
legal conditions of use of the material. Legal
terms often appear because many porno-
graphic Web sites’ entry pages contain a
warning message block.
Most indicative terms occur in the porno-
graphic Web page’s text. We can extract them
from different locations of the correspond-
ing HTML document that might contain
information useful for distinguishing porno-
graphic Web sites. These locations are
• The Web page title
• The warning message block
• Other viewable text in the Web browser
window
• The “description” and “keywords” metadata
• The Web page’s URL and other URLs
embedded in the Web page
• The image tooltip (the text string displayed
when a user points to an object using a
mouse; it usually occurs in the <IMG> tag)
• Graphical text (sometimes graphics or
images contain text that we can extract)
Indicative terms might be displayed or
nondisplayed in the Web browser window.
Displayed terms appear in the Web page title,
warning message block, other viewable text,
and graphical text. Nondisplayed terms are
stored in the URL, the “description” and “key-
words” metadata, and the image tooltip.

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 49

N e u r a l N e t w o r k s

Current Web Content-Filtering Systems

Web-filtering systems are either client- or server-based. A
client-based system performs Web content filtering solely on
the computer where it is installed, without consulting remote
servers about the nature of the Web content that a user tries to
access. A server-based system provides filtering to computers on
the local area network where it is installed. It screens outgoing
Web requests, analyzes incoming Web pages to determine
their content type, and blocks inappropriate material from
reaching the client’s Web browser.
Table A summarizes the features of 10 popular Web-filter-
ing systems (for URLs for these systems, see the “Useful URLs”
sidebar). Only two systems specifically filter pornographic
Web sites. Bold text indicates each system’s main approach.
(For a description of the approaches, see the “Web Content-
Filtering Approaches” sidebar.) No system relies on PICS (Plat-
form for Internet Content Selection) as its main approach. Six
systems rely mainly on URL blocking, and only two systems
mainly use keyword filtering. I-Gear incorporates both URL
blocking and Dynamic Document Review, a proprietary
technique that dynamically looks for matches with keywords.
Only WebChaperone (which is no longer sold or supported)
employs content analysis as its main approach. It uses Internet
Content Recognition Technology (iCRT), a unique mechanism
that dynamically evaluates each Web page before the system
passes the page to a Web browser. iCRT analyzes word count
ratios, page length, page structure, and contextual phrases.
The system then aggregates the results according to the
attributes’ weighting. On the basis of the overall results,
WebChaperone identifies whether the Web page contains
pornographic material.
To gauge the filtering approaches’ accuracy, we evaluated
five representative systems: Cyber Patrol, Cyber Snoop, CYBER-
sitter, SurfWatch, and WebChaperone. We installed each sys-
tem on an individual computer and visited different Web sites
while running the system. We collected the URLs of 200 porno-
graphic and 300 nonpornographic Web pages and used them
for evaluation. To more easily measure each approach’s accu-
racy, we limited each system to only its major approach. So, we
configured Cyber Patrol and SurfWatch to use URL blocking,
Cyber Snoop used keyword filtering, CYBERsitter used context-
based key phrase filtering, and WebChaperone used iCRT.
Table B summarizes our results. We obtained the overall
accuracy by averaging the percentage of correctly classified
Web pages for both pornographic and nonpornographic
pages. For the two systems that employ URL blocking, the
number of incorrectly classified nonpornographic Web pages is
small compared to those using keyword filtering. This shows
that the black list (a list of URLs to block—see the “Web Con-
tent-Filtering Approaches” sidebar) the specialists compiled
accurately excludes URLs of most nonpornographic Web sites,
even if the Web sites contain sexually explicit terms in a non-
pornographic context. However, both systems have fairly high
occurrences of incorrectly classified pornographic Web pages.
This highlights the problem of keeping the black list up to date.
Systems that rely on keyword filtering tend to perform well
on pornographic Web pages, but the percentage of incorrectly
classified nonpornographic Web pages can be high. This high-
lights the keyword approach’s major shortcoming. On the other
hand, WebChaperone, which uses iCRT, achieves the highest
overall accuracy: 91.6 percent. This underlines the effectiveness

Table A. A comparison of 10 popular Web-filtering systems. Bold lettering indicates each system’s main approach.
Content-filtering approach
System Location PICS support URL blocking Keyword filtering Content analysis Filtering domain
Cyber Patrol Client Yes Yes Yes No General
Cyber Snoop Client Yes Yes Yes No General
CYBERsitter Client Yes Yes YesA No General
I-Gear Server Yes Yes YesB No General
Net Nanny Client No Yes Yes No General
SmartFilter Server No Yes No No General
SurfWatch Client Yes Yes Yes No General
WebChaperone Client Yes Yes Yes YesC Pornographic
Websense Server No Yes YesD No General
X-Stop Client Yes Yes YesE No Pornographic

A. Context-based key-phrase filtering technique

B. Employed in Dynamic Document Review and software robots
C. Content analysis carried out using iCRT (Internet Content Recognition Technology)
D. Employed only in Web crawlers
E. Employed in MudCrawler

Table B. The performance of five Web-filtering systems.

Web page
Pornographic (200 total) Nonpornographic (300 total)
System Major approach Correctly classified Incorrectly classified Correctly classified Incorrectly classified Overall accuracy
Cyber Patrol URL blocking 163 (81.5%) 37 (18.5%) 282 (94.0%) 18 (6.0%) 87.75%
SurfWatch URL blocking 171 (85.5%) 29 (14.5%) 287 (95.7%) 13 (4.3%) 90.60%
Cyber Snoop Keyword filtering 187 (93.5%) 13 (6.5%) 247 (82.3%) 53 (17.7%) 87.90%
Cybersitter Context-based key 183 (91.5%) 17 (8.5%) 255 (85.0%) 45 (15.0%) 88.25%
phrase filtering
WebChaperone iCRT 177 (88.5%) 23 (11.5%) 284 (94.7%) 16 (5.3%) 91.60%

50 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

 Useful URLs
Nondisplayed text is contained in HTML Cyber Patrol, www.cyberpatrol.com
Cyber Snoop, www.cyber-snoop.com/index.html
tags (blocks that begin with a “<” and end
CYBERsitter, www.cybersitter.com
with a “>”); displayed text is not. I-Gear, www.symantec.com/sabu/igear
Net Nanny, www.netnanny.com/home/home.asp
Statistics Platform for Internet Content Selection, www.w3.org/PICS
For sample collection, we consider that a SafeSurf, www.safesurf.com
pornographic Web site fulfills any of these SmartFilter, www.smartfilter.com
conditions: SurfWatch, www.surfcontrol.com
Websense, www.Websense.com
• It contains sexually oriented content. X-Stop, www.xstop.com
• It contains erotic stories and text descrip-
tions of sexual acts.
• It shows images of sexual acts, including Table 1. Use of sexually explicit terms in eight locations of 100 samples.
inanimate objects used sexually. Number of unique terms Frequency of occurrence
• It shows erotic full or partial nudity. Location (42 total) (5,673 total)
• It contains sexually violent text or graphics. Web page title 28 (66.67%) 399 (7.03%)
Warning message block 18 (42.86%) 344 (6.06%)
To generally represent such Web sites, we Other viewable text 37 (88.10%) 1,919 (33.83%)
captured 100 different pornographic Web “description” metadata 27 (64.29%) 464 (8.18%)
“keywords” metadata 38 (90.48%) 1,438 (25.35%)
sites and defined their entry pages as sam- URL 29 (69.05%) 581 (10.24%)
ples for data collection. Image tooltip 31 (73.81%) 419 (7.39%)
Graphical text 22 (52.38%) 109 (1.92%)
Page layout format. We found that 86 per-
cent of the entry pages used a single-frame
layout. Among the 14 multiframe pages, 13 Table 2. Use of legal terms in eight locations of 100 samples.
had a two-frame layout. Only one Web page Number of unique terms Frequency of occurrence
had a three-frame layout. Although a single- Location (13 total) (708 total)
frame layout appears prevalent among porno- Web page title 2 (15.38%) 23 (3.25%)
graphic Web sites, our techniques must still Warning message block 13 (100%) 355 (50.14%)
be able to process and analyze multiframe Other viewable text 8 (61.54%) 135 (19.07%)
Web pages. “description” metadata 1 (7.69%) 31 (4.38%)
“keywords” metadata 1 (7.69%) 72 (10.17%)
URL 1 (7.69%) 51 (7.20%)
PICS use. We also found that 89 percent did Image tooltip 2 (15.38%) 39 (5.51%)
not use any PICS support. Of the 11 pages Graphical text 1 (7.69%) 2 (0.28%)
that did, nine used the RSACi content-rating
system; the rest used SafeSurf (see the “Web
Content-Filtering Approaches” sidebar). We These two locations together contribute more not clearly defined, linear or low-order sta-
suspect the low use of PICS among porno- than 59 percent of the total sexually explicit tistical models cannot always describe them.
graphic Web sites is because these sites’ pub- term occurrences, while other locations have So, we use ANNs because they are robust
lishers might not want their sites filtered out. occurrences of approximately six to 10 per- enough to fit a wide range of distributions
The results might not be statistically rep- cent, except graphical text, which contributes accurately and can model any high-degree
resentative of overall PICS use among porno- only 1.92 percent. exponential models.
graphic Web sites. However, they do indicate Table 2 shows that all 13 legal terms appear On the basis of our knowledge of porno-
that implementing PICS support in a Web- in the warning message block; this location graphic-Web-site characteristics, we devel-
filtering system is useful; some pornographic contributes more than 50 percent of the total oped the Intelligent Classification Engine.
sites do have PICS labels embedded in the occurrences of legal terms. Also, more than The engine comprises two major processes:
HTML code’s header section. So, checking 72 percent of the legal term occurrences training and classification (see Figure 1).
PICS labels works best as a supplementary appear in the Web pages’ displayed text. During training, the system learns from sam-
approach for positive identification. ple pornographic and nonpornographic Web
Our ANN approach pages to form a knowledge base of the ANN
Indicative terms. We compiled a list of 55 We could use statistical methods to clas- models. The system then classifies incoming
indicative terms comprising 42 sexually sify Web pages according to the statistical Web pages according to their content.
explicit terms and 13 legal terms. Tables 1 occurrence of certain features. Such methods
and 2 summarize the distribution of these include k-nearest neighbor classification,5 lin- Training
terms in eight locations. As Table 1 shows, ear least-squares fit,6 linear discriminant The Intelligent Classification Engine
more than 90 percent of the terms appear in analysis,7 and naïve Bayes probabilistic clas- extracts the Web page’s representative fea-
the “keywords” metadata, while approximately sification.8 However, because real-world data tures as inputs for training the ANN. Train-
88 percent appear in “Other viewable text.” such as we’re using tend to be noisy and are ing’s objective is to let the ANN configure

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 51

N e u r a l N e t w o r k s

tive terms. So, excluding the URLs will not

Training
compromise the Web page’s relevance to
WWW pornography and will help minimize the sys-
exemplar set
tem’s computational cost.
Feature extraction, then, parses a Web
Training Web pages Online Web pages page to identify the relevant information that
indicates the page’s nature or characteristics.
Engine database
The engine separates the extracted informa-
Feature Feature tion into four types:
extraction extraction
Selected content Selected content • Web page title
Indicative
Preprocessing term Preprocessing • Displayed contents, including the warn-
dictionary ing message and other viewable contents
Terms' occurrence Terms' occurrence • “description” and “keywords” metacontents (the
information contained in the metadata
Transformation Transformation Transformation fields)
weights
Web page vector Web page vector • Image tooltip

Neural Neural Neural Preprocessing. This step converts the raw

network network network
models
New weights Training Activated cluster
results
NN model Cluster-to-
Categorization
generation category
mapping Cluster category
Category Metacontent
assignment checking
Classification
results
Training Classification
Figure 1. The Intelligent Classification Engine.
itself and adjust its weight parameters preprocessing, transformation, ANN model
according to the training exemplars, to facil- generation, and category assignment.
itate generalization beyond the training sam-
ples. This requires a large set of training Feature extraction. We use frequencies of
exemplars to obtain an ANN that produces occurrence of indicative terms in a Web page
statistically accurate results. to judge its relevance to pornography. Our
In our case, the training exemplar is a vec- analysis of pornographic Web pages indicates
tor representing the sample Web page. that indicative terms most likely appear in
Because our objective is to distinguish porno-
graphic Web pages from nonpornographic • Displayed contents consisting of the Web
Web pages, we have collected a total of 3,777 page title, the warning message block, and
nonpornographic Web pages and 1,009 other viewable text
pornographic Web pages for training. To • Nondisplayed contents including the
extensively cover a variety of Web pages with “description” and “keywords” metadata, image
different kinds of content nature, we gath- tooltips, and URLs
ered the nonpornographic Web pages from
13 categories of the Yahoo! search engine. However, we excluded URLs because of
These include Arts & Humanities, Business the difficulties of identifying indicative terms
& Economy, Computers & Internet, Educa- in a URL address. Because words in a URL
tion, Entertainment, Government, Health, are concatenated and are not separated by
News & Media, Recreation & Sports, Refer- white spaces, extracting words or terms from
ence, Science, Social Science, and Society it will be difficult. Furthermore, occurrences
& Culture. The result was a training set of of indicative terms in the URLs in porno-
4,786 exemplars. graphic Web pages contribute only a small
Training has five steps: feature extraction, percentage to the total occurrences of indica-
text extracted from a Web page into numeric
data representing the frequencies of occur-
rence of indicative terms. Figure 2 shows
this step, which has two phases: tokeniza-
tion and indicative-term identification and
counting.
The extracted text must first be parsed into
a list of single words. Because we initially
focus on English-language Web sites, the
tokenization algorithm treats white spaces
and punctuation marks as separators between
words. It converts each extracted word into
lower case before inserting it into a list. To
facilitate identifying phrasal terms, this phase
preserves the word sequence in the list as it
was in the Web page’s raw text. Because fea-
ture extraction produced four types of con-
tents for each Web page, tokenization pro-
duces four word lists: one each for the Web
page title, the displayed contents, the “descrip-
tion” and “keywords” metacontents, and the
image tooltip. Because each list has a differ-
ent degree of relation to the nature of the Web
page, the lists will carry different weights
when training the ANN.
The engine then feeds the four lists into
indicative-term identification and counting.
This phase uses an indicative-term dictionary
to identify the indicative terms in the four
lists. (We manually compiled this dictionary
according to the analysis we described in the
section “Know the enemy.”) It collects the
occurrences for each set of indicative terms
in the dictionary. Because there are 55 dis-
tinctive sets of indicative terms, the system
obtains 55 frequencies of occurrence. Also,
the system counts the total number of indica-
tive-term occurrences and the total number
of words in each list.

52 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Transformation. This step converts the group
of numbers resulting from preprocessing into
the Web page vector to feed into the ANN.
Transformation involves three phases: vec-
tor encoding, weight multiplication, and nor-
malization.
Vector encoding uses the data from pre-
processing to form the basic vector. It has a
dimension of 61 and consists of
• The 55 frequencies of occurrence of
indicative terms (55 elements)
• The total number of indicative term occur-
rences in each of the four lists (four ele-
ments), which represents the relative
importance of the four locations in the
Web page
• The total number of an indicative term’s
occurrences in the four lists (one element),
which represents the term’s degree of rela-
tion to the nature of the Web page
• The total number of words in the four lists,
which indicates the total number of words
in the Web page (one element) and repre-
sents the amount of Web page contents
These elements form the vector representing a
Web page owing to their significance in aiding
Web page content identification. The system
uses the last two elements to train the ANN on
the relation between a Web page’s total number
of indicative terms and total number of words.
Weight multiplication weights the basic
vector according to the 61 elements’ relative
indicativeness. An element’s indicativeness
is the degree of indication it contributes
toward a specific type of content—in this
case, pornography. For example, “xxx” has a
higher relative indicativeness than “babe”
does, while indicative terms in the displayed
contents have a lower relative indicativeness
than those in the metacontents.
Set initially according to the relative
indicativeness, the weights are adjusted
before each ANN training session. So, the
system also treats them as parameters for
fine-tuning the network’s performance to
improve accuracy. It stores the 61 elements’
weights in a weight database with the weight
entry’s position corresponding to the index
of the respective elements.
The engine then normalizes the weighted
vector according to a common Euclidean
length, giving the final Web page vector to
feed to the ANN. This normalization improves
the numerical condition of the problem the
ANN will tackle and makes the training
process better behaved.
SEPTEMBER/OCTOBER 2002
Web page title Displayed contents Metacontents Image tooltip
Preprocessing
Tokenization
Title Displayed content Metacontent Image tooltip
word list word list word list word list
Indicative Indicative term
term identification and counting
dictionary
Frequency of occurrence Total indicative term occurrences Total words in
of 55 indicative terms in each of four lists each of four lists
Figure 2. Preprocessing consists of tokenization, which locates individual words, and
indicative-term identification and counting.
Figure 3 illustrates an example of trans- parameters: the number of training iterations,
formation. the initial and final learning rates, and the ini-
tial neighborhood size and its decrement
ANN model generation. This step trains the interval. For good statistical accuracy, the
ANN for online classification. Two ANNs that number of iterations should be at least 500
have proven effective for document classifi- times larger than the number of output neu-
cation are Kohonen’s Self-Organizing Maps9 rons.9 Because we have 49 neurons, we train
(KSOM) and the Fuzzy Adaptive Resonance the network for 24,500 iterations. In our
Theory10 (Fuzzy ART).11,12 So, we evaluated implementation of the KSOM network, the
them for Web content classification. learning rate linearly decreases in each iter-
Because each vector contains 61 elements, ation from an initial value of 0.6 toward the
the KSOM network requires 61 inputs. We final learning rate of 0.01. The initial neigh-
determined the output neuron array’s dimen- borhood size is set to five and decreases once
sion after conducting a series of training every 4,000 iterations.
experiments on the ANN with several differ- Before a training session begins, we ini-
ent dimensions. We found that if the dimen- tialize the network weights with random real
sion is smaller than 7 × 7, some generated numbers between 0 and 1. Each training iter-
clusters contain a large mixture of porno- ation feeds the whole set of training exem-
graphic and nonpornographic Web pages. plars to the network. The training session
However, output neuron arrays that are larger determines the winning neuron by finding
than 7 × 7 do not significantly improve accu- the response that gives the smallest Euclidean
racy. Also, training an ANN with a larger distance between the exemplar and the
number of output neurons will take much weight vector. Only the weights associated
more time. So, we consider a 7 × 7 output with the winning neuron and its neighbor-
neuron array the best compromise between hood are adjusted to adapt to the input pat-
maximizing the classification distinctiveness tern. When the training session successfully
and minimizing training complexity. completes, the engine saves the network’s
Figure 4 shows the KSOM training algo- weights into the ANN model.
rithm. Before training, we must set several When a training session ends, clusters of
computer.org/intelligent 53
N e u r a l N e t w o r k s
Web page preprocessing

Index Web page data Frequency

Indicative term
1 18 or older | 18 or over |... 1
2 18 years old | 18 year old |... 1
3 of age | of legal age 0
4 adult links | adult link 0
. . .
. . .
. . .
55 xxx 3
Total indicative terms
56 In image tooltip word list 1
57 In displayed content word list 14
58 In metacontent word list 8
59 In title word list 1
60 Sum of all indicative terms 24
61 Sum of all words 236

Vector encoding

Index 1 2 3 4 ... 55 56 57 58 59 60 61
Basic vector 1 1 0 0 ... 3 1 14 8 1 24 236
Weight multiplication

Weighted vector 1 1 0 0 ... 12 3.5 14 32 3 24 236

Normalization

Normalized vector 0.0042 0.0042 0 0 ... 0.0499 0.0146 0.0582 0.1331 0.0125 0.0998 0.9775

Figure 3. An example of transformation. Vector encoding is the result of preprocessing that forms the 61-element basic vector.
Weight multiplication then weighs the elements according to their relative indicativeness. The weighted vector is then normalized
before being fed to the ANN.

Web pages are formed at the array of output

Step 1: Initialize the weight vector of all output neurons. neurons. These clusters might contain dif-
Step 2: Determine the output winning neuron m by searching for the shortest normalized ferent proportions of pornographic and non-
Euclidean distance between the input vector and each output neuron’s weight vector: pornographic Web pages. Category assign-
ment will classify each cluster.
X − Wm = min X − Wj ,
j =1KM The Fuzzy ART network will have 61
nodes at its F0 layer. Owing to complement
where X is the input vector,
Wj is the weight vector of output neuron j, and coding, each F0 node will feed data to two
M is the total number of output neurons. F1 nodes. So, the F1 layer has 122 (61 × 2)
Step 3: Let Nm(t) denote a set of indices corresponding to a neighborhood size of the cur- nodes. Because we wish to compare Fuzzy
rent winner neuron m. Slowly decrease the neighborhood size during the training ART to the KSOM network, the network
session. Update the weights of the weight vector associated with the winner neu-
should generate as close to 49 (7 × 7) clusters
ron m and its neighboring neurons:
as possible. To meet this requirement, the
∆Wj (t ) = α(t)[X(t ) − Wj (t ) ] for j ∈ Nm(t ), engine constructs 60 F2 nodes.
Next, we need to select values for three
where α is a positive-valued learning factor, α ∈ [0, 1]. Slowly decrease α with Fuzzy ART network parameters: the choice
each training iteration.
So, the new weight vector is given by parameter α, the vigilance parameter ρ, and
the learning parameter β. α affects the inputs
Wj (t + 1) = Wj (t ) + α(t ) [X(t ) − Wj (t ) ] for j ∈ Nm(t ) .

Repeat steps 2 and 3 for every exemplar in the training set for a user-defined number of iterations.
Figure 4. The KSOM (Kohonen’s Self-
Organizing Maps) training algorithm.

54 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

 Step 1: Initialize the weights of all weight vectors to 1, and set all F2 layer nodes to
“uncommitted.”
Step 2: Apply complement coding to the M-dimensional input vector. The resultant com-
produced at each node of the F2 layer accord- plement-coded vector I is of 2M-dimensions:
ing to all the nodes of the F1 layer. To mini-
mize the network training time, α must have ( ) (
I = a,aC = a1,K,aM ,a1C ,K,aM
C
),
a small value.10 We chose α = 0.1.
ρ indicates the threshold of proximity to a C
where ai = 1 − ai ,for i ∈ [1, M].
cluster that an input vector must fulfill before
a desirable match occurs. The choice of ρ Step 3: Compute the choice function value for every node of the F2 layer. For the com-
plement-coded vector I and node j of the F2 layer, the choice function Tj is
affects the number of clusters that the Fuzzy
ART network generates. In this case, ρ I ∧W j
Tj (I) = ,
should have a value that lets the network gen- α + Wj
erate approximately 49 clusters. After exper-
imenting with the network several times, we
where Wj is the weight vector of node j, and
chose ρ = 0.66. α is the choice parameter.
β manipulates the adjustment of weight The fuzzy AND operator ∧ is defined by (P ∧ Q)i = min(pi, qi), and the norm | | is
vector W J, where node J is the chosen defined as
cluster that fulfills the vigilance match. We M
use the fast-commit, slow-recode updating P ≡ ∑ pi .
scheme. More specifically, β = 1 when the i =1
chosen cluster is an uncommitted F2 node,
while β = 0.5 for a committed F2 node. Step 4: Find the node J of the F2 layer that gives the largest choice function value:
Unlike the case of KSOM, we do not need to
Tj = max{Tj : J = 1 … N } .
decrease the learning parameter as training
progresses. The output vector Y of layer F2 is thus given by yJ =1 and yj = 0 for j ≠ J.
Figure 5 shows the training algorithm for Step 5: Determine if resonance occurs by checking if the chosen node J meets the vigi-
the Fuzzy ART network. Before the training lance threshold:
session begins, the engine initializes the I ∧ WJ
weights in all weight vectors to 1 and marks ≥ρ ,
I
all the F2 layer nodes as uncommitted. Once
the training session commences, the engine
feeds the training exemplars to the network. where ρ is the vigilance parameter.
If resonance occurs, update the weight vector WJ; the new WJ is given by
This process will carry on until all weight
vectors become static. The engine collects
the network parameters, weights of all ( ) (
WJnew = β I ∧ WJold + 1 − β WJold ) .
weight vectors, and committed status of all
Otherwise, set the value of the choice function TJ to 0.
F2 nodes and stores them as the ANN model Repeat steps 4 and 5 until a chosen node meets the vigilance threshold.
for our Fuzzy ART implementation. For every input pattern in the training set, perform steps 2 to 5.
Similarly to KSOM training, the Fuzzy Repeat the whole process until the weights in all weight vectors become unchanged.
ART training session forms Web page clus-
ters that category assignment will classify. Figure 5. The Fuzzy ART (Fuzzy Adaptive Resonance Theory) training algorithm.

Category assignment. After the ANN gener- Table 3. Thresholds for category assignment.
ates the clusters, we still need to determine Proportion of Web pages in clusters
each cluster’s nature. This is because a clus-
Cluster category Pornographic Nonpornographic
ter defines a group of Web pages with simi-
lar features and characteristics but does not Pornographic [70%, 100%] [0%, 30%)
classify those similarities. Nonpornographic [0%, 30%) [70%, 100%]
The engine assigns each cluster to one of Unascertained [30%, 70%) [30%, 70%]
three categories: pornographic, nonporno-
graphic, or unascertained. Unascertained Table 4. Training efficiency for the Kohonen’s Self-Organizing Maps and the Fuzzy
clusters contain a fair mixture of porno- Adaptive Resonance Theory networks.
graphic and nonpornographic Web pages. Results
Table 3 summarizes the thresholds for cate-
gory assignment. For example, if a cluster Attribute KSOM Fuzzy ART
contains 70 percent of Web pages labeled as Number of inputs 61 61
“pornographic,” we map the cluster to the Number of output neurons 49 47
Number of iterations 24,500 93
“pornographic” category.
Number of training exemplars 4,786 4,786
The system records the results in a cluster- Total processing time 37 hrs., 43 min., 23 sec. 47 sec.
to-category mapping database. Each database

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 55

N e u r a l N e t w o r k s

Table 5. Training accuracy.

ANN type Web page Correctly classified Incorrectly classified Unascertained Total
KSOM Pornographic 911 (90.3%) 27 (2.7%) 71 (7.0%) 1,009
Nonpornographic 3,529 (93.4%) 56 (1.5%) 192 (5.1%) 3,777
Total 4,440 (92.8%) 83 (1.7%) 263 (5.5%) 4,786
Fuzzy ART Pornographic 780 (77.3%) 100 (9.9%) 129 (12.8%) 1,009
Nonpornographic 3,429 (90.8%) 88 (2.3%) 260 (6.9%) 3,777
Total 4,209 (87.9%) 188 (3.9%) 389 (8.1%) 4,786

Table 6. Classification accuracy. The “Meta” column indicates the number of Web pages that metacontent checking classified.
Correctly classified Incorrectly classified
ANN type Web page ANN Meta ANN Meta Unascertained Total
KSOM Pornographic 499 9 23 0 4 535
Nonpornographic 496 1 5 2 19 523
Total 1,005 (95.0%) 30 (2.8%) 23 (2.2%) 1,058
Fuzzy ART Pornographic 428 32 47 0 28 535
Nonpornographic 475 8 7 9 24 523
Total 943 (89.1%) 63 (6.0%) 52 (4.9%) 1,058

entry has a unique ID identifying one of the
generated clusters and its category.
Classification
This process uses the trained ANN to clas-
sify incoming Web pages; it outputs one of
the three predefined categories. Similar to the
training process, it also performs feature
extraction, preprocessing, and transforma-
tion for each incoming Web page. After these
steps have generated the Web page vector,
the system feeds the vector into the ANN to
Figure 6. The engine misclassified this page as pornographic because it contains
sexually explicit terms in its displayed contents and the “description” and “keywords”
metacontents.
produce an activated cluster. The catego-
rization step uses the cluster-to-category
mapping database to determine the activated
cluster’s category, which it uses to classify
the corresponding Web page.
To further reduce the number of unascer-
tained Web pages, we introduce a postpro-
cessing step called metacontent checking.
This step applies keyword filtering to the
“description” and “keywords” metacontents in the
HTML header of the unascertained Web
pages. This mechanism is effective because
the metacontents include terms directly
related to the associated Web pages’ subject.
For keywords, we use the terms in the indica-
tive terms dictionary. If the filtering finds at
least one indicative term in the metacontents,
the engine classifies the associated Web page
as pornographic. Otherwise, it identifies the
page as nonpornographic. If the engine can-
not find the metacontents or they do not exist,
the page remains unascertained.
Performance evaluation
For training, we used the 4,786 Web pages
(93,578,232 bytes) mentioned in the “Train-
ing” section. First, we measured the pro-
cessing time required for the three pre-ANN
steps (feature extraction, preprocessing, and
transformation) for both training and classi-
fication. The three steps took 167 seconds to
process all Web pages (an average of 35 ms
per page).
Next, we measured the training efficiency
and accuracy for KSOM and Fuzzy ART.
Tables 4 and 5 summarize the results.
Although KSOM requires much longer train-
ing time, it produces better training accuracy
and gives a smaller set of unascertained Web
pages.
Finally, we measured both networks’ clas-
sification accuracy. We compiled a testing
exemplar set with 535 pornographic Web
pages and 523 nonpornographic Web pages.
Table 6 summarizes the results. The “Meta”
column indicates the number of Web pages
that metacontent checking classified.
From these results, we conclude that the
KSOM network performed much better than

56 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Fuzzy ART. Overall, the ANNs effectively
generalized information for solving new
cases. Furthermore, a simple technique such
as metacontent checking was also beneficial.
To help us improve the engine, we also
analyzed cases of misclassifications. For
example, the engine misclassified the Web
page in Figure 6 as pornographic. The main
reason for the misclassification is that the
Web page contains many sexually explicit
terms in its displayed contents and the “descrip-
tion” and “keywords” metacontents.
Another common reason for misclassifi-
cation is the insufficient text information in
some Web pages that contain mainly image
information. If a page’s metacontents do not
contain any information, and the page does
not display any warning messages as text, the
system will misclassify it.
T h e A u t h o r s
Pui Y. Lee is a graduate student in Nanyang Technological University’s
School of Computer Engineering. His research interests are neural networks,
Web page classification, intelligent filtering, and Internet technology. He
received his BASc (Hons.) in computer engineering from NTU. Contact him
at the School of Computer Eng., Nanyang Technological Univ., Blk N4,
#02A-32, Nanyang Ave, Singapore 639798; a3397245@ntu.edu.sg.
Siu C. Hui is an associate professor in Nanyang Technological University’s
School of Computer Engineering. His research interests include data min-
ing, Internet technology, and multimedia systems. He received his BSc in
mathematics and his DPhil in computer science, both from the University of
Sussex. He is a member of the IEEE and ACM. Contact him at the School of
Computer Eng., Nanyang Technological Univ., Blk N4, #02A-32, Nanyang
Ave., Singapore 639798, Republic of Singapore; asschui@ntu.edu.sg; www.
ntu.edu.sg/sce.
Alvis Cheuk M. Fong is a faculty member in Massey University’s Institute
of Information and Mathematical Sciences. His research interests include
various aspects of Internet technology, information theory, and video and
image signal processing. He received his BEng and MSc from Imperial Col-
lege, London, and his PhD from the University of Auckland. He is a mem-
ber of the IEEE and IEE, and is a Chartered Engineer. Contact him at the Inst.

C ompared to Web-filtering systems we

have surveyed (see the “Current Web
Content-Filtering Systems” sidebar), our
Intelligent Classification Engine based on the
KSOM network model is clearly more accu-
rate. Our results indicate that the engine can
automate the maintenance of a URL black
list (a list of URLs to block—see the “Web
Content-Filtering Approaches” sidebar) for
fast, effective online filtering. Also, auto-
mated content classification minimizes the
need to manually examine Web pages.
Our research is advancing in two direc-
tions. First, we are developing heuristics to
complement machine intelligence for better
classification accuracy. A particular focus is
Web pages that do not contain much text. We
are applying pattern recognition techniques
for understanding graphical information.
Second, we are extending the work to cover
multilingual, XML, and nonpornographic
but objectionable (for example, violence- and
drug-related) Web pages. Because we have
established the framework, the changes
require only the extraction of exemplars that
characterize such Web content.
of Information and Mathematical Sciences, Massey Univ., Albany Campus,
Private Bag 102-904, North Shore Mail Centre, Auckland, New Zealand;
a.c.fong@massey.ac.nz; www.massey.ac.nz/~acfong.
For more information on this or any other computing topic, please visit our Digital Library at
http://computer.org/publications/dlib.
References 7. G.J. Koehler and S.S. Erenguc, “Minimizing
Misclassifications in Linear Discriminant
1. G. Salton, Automatic Text Processing, Addi- Analysis,” Decision Sciences, vol. 21, no. 1,
1990, pp. 63–85.
son-Wesley, Boston, 1989.
2. R.P. Lippmann, “An Introduction to Com- 8. A. McCallum and K. Nigam, “A Comparison
puting with Neural Networks,” IEEE ASSP of Event Models for Naïve Bayes Text Clas-
Magazine, vol. 4, no. 2, Apr. 1987, pp. 4–22. sification,” AAAI-98 Workshop Learning for
Text Categorization, AAAI Press, Menlo
Park, Calif., 1998, pp. 41–48.
3. M. Pastore, “Search Engines, Browsers Still
Confusing Many Web Users,” CyberAtlas,
INT Media Group, Darien, Conn., 14 Feb. 9. T. Kohonen, Self-Organizing Maps, Springer-
2001, http://cyberatlas.internet.com/big_picture/ Verlag, Berlin, 1995.
applications/article/0,,1301_588851,00.html.
10. G.A. Carpenter, S. Grossberg, and D.B.
4. The Effects of Pornography and Sexual Mes- Rosen, “Fuzzy ART: Fast Stable Learning and
Categorization of Analog Patterns by an
sages, National Coalition for the Protection
of Children & Families, Cincinnati, Ohio, Adaptive Resonance System,” Neural Net-
www.nationalcoalition.org/pornharm. works, vol. 4, no. 6, July 1991, pp. 759–771.
phtml?ID=102.
11. G. Troina and N. Walker, “Document Classi-
5. Y. Yang, “An Evaluation of Statistical fication and Searching: A Neural Network
Approach,” ESA Bull., no. 87,Aug. 1996; http://
Approaches to Text Categorization,” Infor-
mation Retrieval, vol. 1, nos. 1–2, Apr. 1999, esapub.esrin.esa.it/bulletin/bullet87/troina87.
pp. 69–90. htm.

Acknowledgments
The list of products surveyed in this article is
not exhaustive. We are not related to any of the
vendors, and we do not endorse any of these
products.
6. Y. Yang and C.G. Chute, “An Application of
Least Squares Fit Mapping to Text Informa-
tion Retrieval,” Proc. 16th Ann. Int’l ACM
SIGIR Conf. Research and Development in
Information Retrieval (SIGIR 93), ACM
Press, New York, 1993, pp. 281–290.
12. D. Roussinov and H. Chen, “A Scalable Self-
Organizing Map Algorithm for Textual Clas-
sification: A Neural Network Approach to
Thesaurus Generation,” Communication and
Cognition—Artificial Intelligence, vol. 15,
nos. 1–2, Spring 1998, pp. 81–111.

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 57