We introduce a methodology for identifying approximately which words attract the users attention when they are visiting the pages in a web site. These words are called “web site keywords” and can be used for creating further web page text contents related with an specific topic. Through the utilization of the correct words, we can help to the users to find what they are looking for. By applying a clustering algorithm, and under the assumption that there is a correlation between the time spent in a page and the user interest, we segment the users by navigation behavior and content preferences. Next we identify the web site keywords. This methodology was applied on data originated in a real web site, showing its effectiveness.
For many companies and/or institutions, it is no longer
sufficient to have a web site and high quality products or
services. What often makes the difference between e-business
success and failure is the potential of the respective web site
to attract and retain users. This potential depends on the
site content, design, and technical aspects, such as the time
to download the pages from the web site to the user’s web
browser, among others. In terms of content, the words used
in the free text of a web site pages are very important, as the
majority of the users perform term-base queries in a search
engine to find information on the Web. These queries are
formed by keywords, i.e., a word or a set of words [
The suitable utilization of words in the web page
improves user appeal, helps effective information search,
while attracting new users and retaining current users by
continuous updating of page text content. So the
challenge is to identify which words are important for users.
Most keywords are calculated from “most frequently used
words”. Some commercial tools1 help identify target
keywords that customers are likely to use while web searching
[
By identifying the most relevant words in the sites pages, from the user point of view, improvements can be performed in the entire web site. For instance, the site could be restructured putting a new hyperlink related with the keyword and of course the text content could be modify by using the keywords related with an specific topic to enrich the free text in a web page.
In this paper a methodology for analyzing the user browsing behavior and text preferences is introduced, through the application of web mining algorithms on data originated in the Web, also called web data, specify web log registers and web site text content.
The methodology aims to identify approximately which
words attract the users attention when they are visiting the
pages in a web site. These words are called “web site
keywords” [
This paper is organized as follow. Section 2 introduces a short revision about the related work. The preparation process for transforming the web data in feature vectors to be used as input in web mining algorithm, is shown in section 3. In section 4, the methodology for identifying the web site keywords is explained and applied in section 5. Finally, the section 6 shows the main conclusions of this paper. 2
When a user visit a web site, data about the pages visited are stored in the web log files. Then it is direct to know which pages were visited and which not, and inclusive the time spent by a user in each one of them. Because usually the pages contain information about an specific topic, it is possible to know approximately the users information preferences. In that sense the interaction between user and site is like an electronic inquest, given us the necessary data for 1see e.g. http://www.goodkeywords.com/ analyzing the user content preferences in a particular web site.
The challenge for analyzing the user text preferences in the site is double. First the amount of web log registers usually come be huge, and an important part of them with irrelevant information about the user browsing behavior in the site. Second, the free text inside of the web pages is commonly plain, i.e., without additional information that allow us to know directly which words attract the user attention.
In this section the main approaches to analyze web data for extracting significant patterns related with the users text preferences in a web site are reviewed. 2.1
Web mining techniques emerged as a result of the
application of data mining theory to pattern discovery from web
data [
The following common terminology is used to define the different types of web data: • Content. The web page content, i.e., pictures, free text, sounds, etc. • Structure. Data that shows the internal web page structure. In general, they have HTML or XML tags, some of which contain information about hyperlink connections with other web pages. • Usage. Data that describes visitor preferences while browsing in a web site. It is possible to find such data inside web log files. • User profile. A collection of information about a user: personal information (name, age, etc.), usage information (e.g. visited pages) and interest.
With the above definitions and depending of the web data to be processed, web mining techniques can be grouped in three areas: Web Content Mining (WCM), Web Structure Mining (WSM), and Web Usage Mining (WUM). 2.2
The goal is to automatically construct summaries of a
natural-language document [
In text summarization research, the three major
approaches are [
The first approach consists in selecting a single
paragraph of a text segment [
In the second approach, the most interesting phrases or
key-phrases are extracted and assembled in a single text
[
The key-text components are parts of an entire document, for instance a paragraph, phrase and up to a word, that contain significant information about a particular topic, from the web site user point of view. The identification of these components can be useful for improving the web site text content.
Usually, the keywords in a web site are correlated with
“most frequently used words”. In [
Another method, in [
Finally, instead of analyzing words, in [
Of all available web data, the most relevant for the
analysis of user browsing behavior and preferences, are the web
log registers and the web pages [
The process of segmenting the visitors activities into
individual visitor sessions is called sessionization [
Many authors [
The session reconstruction aims to find the real user sessions, i.e., which pages were visited by a physical human being. In that sense, whatever the chosen strategy to discover real sessions, it must satisfy two essential criteria: the activities performed by a real person can be grouped together and the activities that belong to the same visit (others object required for the visited page) also belong to the same group.
There are several techniques for sessionization, which
can be grouped in two major strategies: proactive and
reactive [
Proactive strategies aim to identify the user using
identification methods like cookies and these consist in a piece
of code associated with the web site. When a visitor visits
the site for the first time, a cookie is sent to the browser.
Next, when the page is revisited, the browser shows the
cookie content to the web server, and an automatic
identification takes place. The method has problems from a
technical point of view and also with respect to the
visitor’s privacy. First, if the site is revisited after several hours,
the session will be considered too long, it will actually be
a new session. Secondly, some aspects of the cookies seem
to be incompatible with the principles of data protection of
some communities, like the European Union [
Reactive strategies are noninvasive with respect to privacy and they make use of the information contained in the web log files only and consist in processing the registers to generate a set of reconstructed sessions.
In web site analysis, the general scenario is that the web
sites usually don’t implement identification mechanisms.
The utilization of reactive strategies can be more useful.
They can be classified into two main groups [
There are several methods for comparing the content of two web pages, here considered as the free text inside the web pages. The common process is to match the terms that make up the free text, for instance, by applying a word comparison process. A more complex analysis includes semantic information contained in the free text and involves an approximate term comparing task as well.
Semantic information is easier to extract when
documentation includes additional information about the text
content, e.g., market language tags. Some web pages allow
document comparison by using the structural information
contained in HTML tags, although with restrictions. This
method is used in [
Comparisons are made by a function that returns a
numerical value showing the similarities or differences
between two web pages. This function can be used in the
web mining algorithm to process a set of web pages, which
might belong to a web community or an isolated web site.
The comparison method must consider efficiency criteria
in the web page content processing [
Web pages must be cleaned before transforming them
into vectors, both to reduce the number of words - not all
words have the same weight - and make the process more
efficient. Thus, the process must consider the following types
of words:
• HTML Tags. In general, these must be cleaned.
However, the information contained in each tag can be used
to identify important words in the context of the page.
For instance, the <title> tag marks the web page
central theme, i.e., gives an approximate notion of the
semantic meaning of the word and, is included in the
vector representation of the page.
• Stop words (e.g. pronouns, prepositions, conjunctions,
etc.)
• Word stems. After applying a word suffix removal
process (word stemming [
For vector representation purposes, let R be the total number of different words and Q be the number of pages in the web site. A vectorial representation of the set of pages is a matrix M of size RxQ,
M = (mij ), i = 1, . . . , R and j = 1, . . . , Q, (1) where mij is the weight of word i in page j.
Based on tfxidf-weighting introduced in [
Here, fij is the number of occurrences of word i in page j and ni is the total number of times that the word i appears in the entire web site. Additionally, a words importance is augmented by the identification of special words, which correspond to terms in the web page that are more important than others, for example, marked words (using HTML tags), words used by the user in search of information and, in general, words that imply the desires and the needs of the users. The importance of special words is stored in the array sw of dimension R, where sw(i) represents an additional weight for the ith word.
The array sw allows the vector space model to include ideas about the semantic information contained in the web page text content by the identification of special words.
The common sources of special words are: 1. E-Mails. The offer to send user e-mails to the call center platform. The text sent is a source to identify the i most recurrent words. Let ewi = weT−mEail be the array of words contained in e-mails, which are also present in the web site, where wei−mail is the frequency of the ith word and T E is the total amount of words in the complete set of e-mails. 2. Marked words. Within a web page, there are words with special tags, such as a different font, e.g., italics, or a word belonging to the title. Let mwi = wmiTaMrks be the array of marked words inside web pages, where wmiark is the frequency of the ith word and T M is the total amount of words in the whole web site. 3. Asking words. A bank, for example, has a search engine through which the users can ask for specific subi jects, by introducing key words. Let awi = wTaAsk be the array of words used by the user in the search engine and also contained in the web site, where waisk is the frequency of the ith word and T A is the total amount of words in the complete set. 4. Related web site. Usually a web site belongs within a market segment, in this case the financial institutions market. Then, it is possible to collect web site pages belonging to the other sites in the same market. Let rwi = RwWriwSs be the array with the words used in the market web sites including the web site under study, where wriws is the frequency of the ith word and RW S is the total number of words in all web sites considered.
The final expression swi = ewi + mwi + awi + rwi is the simple sum of the weights described above.
In the vectorial representation, each column in the matrix M is a web page. For instance the kth column mik with i = 1, . . . , R is the “kth” page in the entire set of pages.
WPk = (wp1k , . . . , wpkR) = (m1k, . . . , mRk), k = 1,. . . ,Q , is the vectorial representation of the kth page in the set of pages under analysis.
With the web pages in vectorial representation, it is possible to use a distance measure for comparing text contents. The common distance is the angle cosine calculated as dp(W P i, W P j) =
!kR=1 wpikwpjk "!kR=1(wpik )2"!kR=1(wpjk)2 . (3)
The Eq. (3) allows to compare the content of two web
pages, returning a numerical value between [
Different techniques are applied to analyze web site user behavior ranging from simple web page use statistics to complex web mining algorithms. In the last case, research concentrates on predictions about which page the user will visit next and the information they are looking for.
By using mainly a combination of a WUM and WCM approaches, it is propose to analyze the web user text preferences in a web site and for this way, to identify which words attract the user attention during their navigation in the site.
Prior to the application of a web mining tool, the data related to web user behavior has to be processed to create feature vectors, whose components will depend on the particular implementation of the mining algorithm to be used and the preference patterns to be extracted. 4.1
The majority of the web user behaviour models examine
the sequence of pages visited to create a feature vector that
represents the web user’s browsing profile in a web site [
υ = [(p1, t1) . . . (pn, tn)], where (pi, ti) are the parameters that represent the ith page from a visit and the percentage of time spent on it in the session, respectively. In this expression, pi is the page identifier.
In Definition 2, the user behavior in a web site is characterized by: 1. Page sequence; the sequence of pages visited and registered in the web log files. If the user returns to a page stored in the browser cache, this action may not be registered. 2. Page content; represents page content, which can be free text, images, sounds, etc. For the purposes of this paper, the free text is mainly used to represent the page. 3. Spent time; time spent by the user in each page. From the data, the percentage of time spent in each page during the user session can be directly calculated. 4.2
The aim is to determine the most important words at a given web site for users, by comparing the user text preferences, through the analysis of pages visited and the time spent on each of them [34]. It differs, however, from the previously mentioned approaches, as the exercise is to find the keywords that attract and retain users from the user web usage data available. The expectation is to involve current and past users in a continuous process of keywords determination.
User preferences about web content are identified by
content comparison of pages visited, [
From the user behavior vector (UBV), the most important pages are selected assuming that degree of importance is correlated to the percentage of time spent on each page. The UBV is sorted according to the percentage of total session time spent on each page. Then the ι most important pages, i.e. the first ι pages, are selected.
tor ϑι(υ) = [(ρ1, τ1), . . . , (ρι, τι)], where (ρι, τι) is the component that represents the ιth most important page and the percentage of time spent on it by session.
Let α and β be two UBVs. The proposed similarity measure between the two IPVs is introduced in equation 4 as: st(ϑι(α), ϑι(β)) = 1 ι τkα , τkβ ι k#=1 min{ τkβ τkα } ∗ dp(ρkα, ρkβ ) (4)
The first element in (4) indicates the users interest in the visited pages. If the percentage of time spent by users α and β on the kth page visited is close to each other, the value of the expression min{·, ·} will be near 1. In the extreme opposite case, it will be near 0. The second element in (4) is dp, the distance between pages in the vectorial representation introduced in (3). In (4) the content of the most important pages is multiplied by the percentage of total time spent on each page. This allows pages with similar contents to be distinguished by different user interests. 4.3
A web site keyword is defined as “a word or possibly a
set of words that make a web page more attractive for an
eventual user during his/her visit to the web site” [
In order to find the web site keywords, it is necessary to
select the web pages with text content that is significant for
users. The assumption is that there exists a relation between
the time spent by the user in a page and his/her interest in
its content [
Among these web mining techniques, special attention
should be paid to the clustering algorithms. The assumption
is that, given a set of clusters extracted from data generated
during former user sessions in the web site, it is possible to
extract the users preferences by analyzing the clusters
content. The patterns in each cluster detected would be
sufficient to extrapolate the content that he or she is looking for
[
In each IPV, the page component has a vectorial representation introduced by the Eq. (2). In this equation, an important step is the calculus of the weights consider in the special words array swi. The special words are different of a normal word in the site because they belong to an alternative and related source or they have an additional information showing their importance in the site, for instance and HTML tag emphasized a word .
The clustering algorithm is used for grouping similar IPVs by comparing the time and page components of each vector, being important to use the the similarity measure introduced in the Eq. (4). The results should be a set of clusters whose quality must be checked by using an accept/reject criterion. A simple way is to accept the clusters whose pages share a same main them and in otherwise to reject the cluster. In that point, it is necessary which page in the site is closet with the vectors in the cluster. Because we know the web site pages vectorial representation and using the Eq.(3) we can identify the closet page to a given cluster’s vector and for this way to get the associate pages of the cluster and to review if the pages share a common main theme.
For each accepted cluster and remembering that the centroids contain the pages where the users spent more time during their respective sessions and in vectorial representation the special words have the highest weights, the procedure for identify the web site keywords is to apply a measure, described in the Eq. (5) (geometric mean) to calculate the importance of each word kw[i] = $ι% mip p∈ζ (5) where i = 1, . . . , R, kw is an array containing the weights for each word relative to a given cluster and ζ the set of pages representing this cluster. The web site keywords are the result of sorting kw and to detect the words with highest weights, for instance the ten words. 5
For experimental purposes, the selected web site should be complex with respect to several features: number of visits, periodic updating (preferably monthly in order to study the user reaction to changes) and rich in text content. The web page of a Chilean virtual bank (no physical branches, all transactions undertaken electronically) met these criteria. As noted, the authors signed a non-disclosure agreement with the bank and are not in a position to provide its name.
The main characteristics of the bank web site are the following; presented in Spanish, with 217 static web pages and approximately eight million raw web log registers for the period under consideration, January-March 2003.
The behavior of a user at the bank web site is analysed in two ways. First, by using web log files which contain data about visitor and customer browsing behavior. This data requires prior reconstruction and cleaning before web mining tools are applied. Second, web data is the web site itself, specifically the web page text content - this also needs preprocessing and cleaning. 5.1
Fig 1 shows part of the bank’s web log registers and includes both identified customers and anonymous vistors. Customers access the site through a security connection, using a SSL protocol that allows the storage of an identification value in the authuser parameter in the web log file. Another way of identifying users is by cookies, but sometimes these are disactivated by users in their browsers. In this case it will be necessary to reconstruct the visitor session.
During the session reconstruction process, filters are applied to the web logs registers. In this particular case, only the registers requesting web pages are used to analyze the site specific user behavior. It is also important to clean abnormal sessions, for example web crawlers, as is shown in Fig. 1 line 4, where a robot that belongs to Google is detected.
The raw web logs cover four months of transactions, with approximately eight millions registers. Only registers related to web pages are considered for the session reconstruction and user behavior analysis purposes; information that points to other objects, like pictures, sounds, etc., will be cleaned. 5.2
By applying web page text filters, it was found that the complete web site contains R=2,034 different words to be used in the analysis. Regarding the word weights and the special words specification, the procedure introduced in section 3.2 was used, in order to calculate swi in equation 2. The data sources were: 1. Marked words. Inside the web pages, 743 different words were found after applying the preprocessing and cleaning step. 2. Related web sites. Four web sites were considered, each of them with approximately 300 pages. The total number of different words was 9253, with 1842 of them contained in the web site text content.
After identifying the special words and their respective weights, it is possible to calculate the final weight for each word in the entire web site, by applying the Eq. 2. Then, the vector representation for all the pages in the site is obtained. 5.3
We fixed at 3 the vector’s maximum dimension. Then, a SOFM with 3 input neurons and 32 output neurons was used to find clusters of Important Page Vectors.
Figure 2 shows the neurons positions within the SOFM on the x, y axes . The z axis is the normalized winning frequency of a neuron during training.
Figure 2, shows 8 main clusters which contain the information about the most important web site pages. However, only 5 were accepted. The accept/reject criterion is simple; if the pages in the cluster centroid have the same main theme, then the cluster is accepted - otherwise it is rejected.
The cluster centroids are shown in Table 1. The second column contains the center neurons (winner neuron) of each cluster and represents the most important pages visited.
A final step is required to get the web site keywords; to analyze which words in each cluster has a greater relative importance in the complete web site.
The keywords and their relative importance in each cluster are obtained by applying the equation 5. For example, if the cluster is ζ = {16, 159, 173}, then kw[i] = √3mi16mi159mi173, with i = 1, . . . , R.
Finally, by sorting the kw in descendending order, we can select the k most important words for each cluster, for example k = 5.
We are not able to show the specific keywords because of the confidentiality agreement with the bank. For this reason the words are numbered. Table 2 shows the keywords found by the proposed method.
Table 3 shows a selected group of keywords from all clusters. Keywords on their own, however, do not make much sense. They need a web page context where they could be used as special words, e.g. marked words to emphasize a concept or as link words to other pages.
The specific recommendation is to use the keywords as “words to write” in a web page, i.e., the paragraphs written in the page should include some keywords and some could be linked to other pages.
Further it is possible on the basis of this exercise to make recommendations about the text content. However, to reiterate, keywords do not work separately for they need a context. Reviewing Table 2, for each cluster, the discovered keyword could be used to rewrite a paragraph or an entire page. In addition, it is important to insert keywords to highlight specific concepts.
Keywords can also be used as index words for a search engine,i.e., some could be used to customize the crawler that visits web sites and load pages. Then, when a user is looking for a specific page in a search engine, the probability of getting the web site increases.
Web site keywords are concepts to motivate the users’ interests and make them visit the web site. They are to be judged within their context for as isolated words they may make little sense, since the clusters represents different contexts. The specific recommendation is to use the keywords as “words to write” in a web page.
Web site keywords can also be used as search engine index words, i.e., some of them could be used to customize crawlers that visit web sites and load pages. Then, when a user is looking for a specific page in the search engine, the probability of getting the web site increases.
As each page contains a specific text content, it is possible to associate the web site keyword to the page content; and from this suggest new content for site revision or reconstruction. For example, if the new page version is related to the “credit card”, then the web site keywords “credit, points and promotions” must be designed for the rewritten page text content. 6
When the users visit a web site, there is a correlation between the maximum time spent per session in a page and its free text content. Then we created the “Important Page Vector (IPV)”, which is the basic data structure for storing the pages where the user spent more time during his/her session. By using the IPV as input in a SOFM, we can to identify clusters that contain the user navigation and information preferences.
The cluster accept/reject criterion is simple: if the pages inside in each cluster are related with a similar main topic, then the cluster is accepted, else rejected. Applying this criterion, five cluster were accepted and the patterns contained in each on them were used for extracting the web site keywords.
The web page text contain can be improved by using the web site keywords, and for this way to attract the user attention when he/she is visiting the site. However, it is necessary to remember that these words cannot be used isolated, in fact they need a context, which usually is provided by a human being.
As future work, we will apply the methodology in other web data, for instance the images, in order to identify which elements attract the user attention in the web site.
This work was supported partially for the Millennium Institute on Complex Engineering Systems. [34] J. D. Vela´squez, H. Yasuda, T. Aoki, and R. Weber. Using the kdd process to support the web site reconfiguration. In Procs. IEEE/WIC Int. Conf. on Web Intelligence, pages 511– 515, Halifax, Canada, October 2003. [35] J. D. Vela´squez, H. Yasuda, T. Aoki, and R. Weber. A new similarity measure to understand visitor behavior in a web site. IEICE Transactions on Information and Systems, Special Issues in Information Processing Technology for web utilization, E87-D(2):389–396, February 2004. [36] J. Xiao, Y. Zhang, X. Jia, and T. Li. Measuring similarity of interests for clustering web-users. In ADC ’01: Proceedings of the 12th Australasian conference on Database technologies, pages 107–114, Washington, DC, USA, 2001. IEEE Computer Society. [37] K. Zechner. Fast generation of abstracts from general domain text corpora by extracting relevant sentences. In Procs. Int. Conf. on Computational Linguistics, pages 986–989, 1996.