Web portals contain large amount of information. Users could really benefit from it if personalized presentation is used. For this to be accomplished the website needs to “know” its users. When surfing the Web users leave digital footprints in the form of navigational paths and actions taken. We present a method for adaptive navigation support and link recommendation. The method is based on an analysis of the user's navigational patterns and behavior on the web pages while browsing through a web portal. We extract interesting information from a web portal and then recommend it. Finally, we provide our experience with a recommender system deployed on our faculty's website, which recommends events by means of personalized calendar.
Web portals are being visited by various users pursuing different
goals. However, most websites offer all groups of visitors the
same content. Therefore the visitors are often presented
information in which they have no interest [
While browsing through a web portal some users can discover interesting pages that are hidden deeper in the hierarchy of the portal. If the users with similar goals knew about these pages they could find them interesting, too. Personalized navigation and recommendation based on monitoring user activities and social principles is a viable approach for such cases. In this paper we introduce a method of navigation adaptation and social recommendation of links among users with similar behavior.
People use different means of accessing information on a web
portal. The most common way is to follow hyperlinks, which
accounts for more than half of all possibilities [
User’s habits can be derived from the navigational patterns found
in the sequences of links he uses in a particular web portal. Four
basic navigational patterns (path, loop, ring and spike) were
described in [
We propose a method for adaptive recommendation of interesting links in a particular web portal (which we may or may not own). For a specific user we select links that similar users found interesting. We also recommend links to this user based on his previous surfing sessions. Our recommender system extracts further information from the web pages, which is also shown to the user. The recommended links are shown in special sections added to each web page of the portal.
When deciding which link to recommend we do not consider the content of web pages. We based our recommendations solely on the analysis of user’s behavior. Our method thus does not depend on the language of the website. We are able to analyze interest and patterns on different language versions of the same portal. Our method of adaptive link recommendation works in two steps: 1. 2.
Recommendation of links based on user’s behavior.
In the first step we analyze the sequences of followed links from each user’s session. In these sequences we look for navigational patterns. We use the prevailing pattern to categorize users, as it determines user’s surfing habits on a particular web portal. As an output we get groups of users with similar navigational patterns. In the second step we monitor behavior of users on each visited web page of the portal. From their actions we automatically estimate their interest in that page. We then recommend links to interesting pages among users of each group from the step one.
We find similar users based on comparison of navigational
patterns they follow in a closed web portal. We believe that users
who follow analogous paths should be recommended similar
links. There are four basic navigational patterns according to [
Ring – a sequence that starts and ends in the same node. Loop – a sequence that goes through already visited node.
Spike – a sequence that goes back through the same trail.
In each session a user visits several pages of the web portal. This
session is described by a vector whose elements are links to the
web pages arranged in order they were visited. We consider
a continuous sequence of links to be a session. For this purpose
1:
2:
3:
4:
5:
6:
1:
2:
3:
4:
5:
we use the referrer field of HTTP request message. If the URL of
previously visited page equals referrer value of currently visited
page, we consider the pages to be in the same session.
The process of dividing users into groups is presented in Alg. 1.
Similarity of users is expressed as Pearson correlation coefficient
[
Algorithm 1 Group users according to their similarity. for each user u do find patterns in clickstreams of u put u to group according to prevailing pattern for each group g do for each user u in group g do sort users in group g according to their similarity to u Every user ends up in exactly one group according to the most dominant pattern found in his surfing history. There is one more group for users with no dominant navigational pattern. The order of similar users from one group is unique for each user. Alg. 2 presents the process of recommending links among users.
Algorithm 2 Recommend links for user u by similar users. similar = select top K similar users for each user v in similar do for each page p visited by v and not visited by u do
predict interest of user in page (u, p) recommend top M pages with highest predicted interest Navigational patterns of users have to be of a certain minimal length (so that each sequence of two following pages does not represent a path pattern). After finding similar users to user u we select top K of them to form a recommendation group. The groups change according to new browsing sessions in which the users can behave differently. This reflects the evolution of user’s behavior in time. However, at each time the user belongs to exactly one group according to the most dominant pattern in browsing history.
In order to recommend links to a particular user we need to evaluate the interests of the users in his recommendation group. We can recommend pages which other users liked. To determine user’s interest in a particular web page we observe actions he makes on this page. These include time spent on a web page, number of scrolling events that occur and number of times he copies text into the clipboard. We chose these interest indicators because their tracking is platform independent.
Our method is based on the comparison of current user’s behavior with the behavior of other users. We compare the values of time and scrolling with values from other people who visited the same page. If the value for the current user is more than X % higher than the average, we consider it as a sign of positive interest in the page. In contrast, when it is more than X % lower than the average we consider it as a sign of negative interest. When the value is around average (± X %) it is a sign of neutral interest. This way we can also consider other interest indicating actions. The exact value of X depends on the calibration for selected domain; in our experiments we used the value of 20 %. When no behavioral data for a particular web page is available we cannot estimate the user’s interest. This is a problem with newly added pages as well as with pages visited for the first time. We estimate the actual value of user’s interest in each page he visits according to Figure 1. We increase this value by 0.1 when the user also copied text into clipboard; otherwise we decrease it by 0.1. Resulting interest is in the interval <0,1> with 0 meaning no interest and 1 meaning total interest in the visited web page.
We recommend web pages by predicting user’s interest in yet
unseen pages using collaborative filtering method. We compute
the predicted value of interest like this [
N u 1 (ru,i ru )
N u 1 Sa,u
Sa,u
where pa,i means prediction of interest of user a in page i, ra is the
average interest of user a in all visited pages, ru,i is the interest of
user u in visited page i, Sa,u is the value of Pearson correlation
coefficient [
To evaluate the proposed method for user’s interest estimation we
developed software tools which support adaptive navigation by
recommending information extracted from potentially interesting
web pages to guests of particular web portal. We experimented
with the web portal of our faculty (www.fiit.stuba.sk).
We proposed client-server architecture with an adaptive proxy [
Adaptive Proxy Server
WebImp Plugin add personalized content
Behavior Plugin add tracking JavaScript store actions response request web pages Domain model User model pages, events anSalpyyzIemppages
create events behavior, analyzed pages crawl recommendations
AdaptiveImp find navigational patterns find similar users estimate user’s interest create recommendations
Website Our plug-in to the adaptive proxy (WebImp) modifies the web page by adding special sections with recommended links. One of those sections is personalized calendar. Many web pages on the web portal of our faculty inform about an upcoming event. We automatically extract dates from these pages and create events. Using proposed method we determine user’s interest in such page. Then, if the interest is positive, we add the event to user’s calendar. This way we also recommend events among users. We monitor the portal and capture every change in text of a web page (this could be for example a change in time and place of some event). Every changed page is marked as news and added to a special news section. We also recommend other potentially interesting links which are neither events nor news. Figure 3 shows part of a web page enhanced with personalized sections.
We provided a series of experiments on our faculty website. Actions of 24 users on a modified website were monitored for 3 weeks. We compared our calculations with their explicit feedback. Results indicate that time actively spent on a web page is the best interest indicator. Scrolling proved to indicate positive interest as well. However, when the user does not use scrolling, it does not always mean he is not interested in the page. The accuracy of our interest estimation method was 62 %.
The sections with recommended links – especially calendar – were attractive (according to answers from questionnaire) and the users found 55 % of recommended links and events interesting. Some users were not satisfied with the recommendations. The problem was that they visited the website for the first time. Hence their user model was empty and we could not provide suitable recommendations. This is a common problem with recommender systems and new users. We tried to overcome it by recommending the most interesting events (links) according to behavior of all users. However, this is not always a suitable solution.
We have presented a method for adaptive recommendation of interesting links. Our approach is based on collaborative filtering, which has a potential to be used in unusual ways. We presented one of them when considering data about user’s actions instead of content of pages. This way we are able to predict user’s interest for unvisited pages. Our method achieves solid results and can be further improved in a recommender system which will combine content analysis with behavior, which is our plan for future work. In this paper we presented a useful application of our method by creating personalized calendar of events on our faculty’s website. Using this method we can also personalize other sections of a web page as well. In our opinion recommender systems should bring added value to users by doing further analysis of the domain which is being adapted. On the web they should recommend particular objects (e.g. events) instead of simply listing potentially interesting links to web pages. In order to accomplish this we need to use more text processing algorithms in our recommender systems so that they “understand” the meaning of text on the web. We ran up against a problem with incorporating the sections with personalized content into a website. In order to do this we need to know the semantics of the website’s structure. This is also useful during page analysis and content extraction. We consider the special tags in HTML5 (e.g. nav, footer) to be insufficient so we came up with a descriptive XML file which pairs HTML tags and their IDs with their predefined meaning (left menu, right menu, etc.). This way our recommender system understands the structure of a website and can alter some sections. We think that there is a need for further development of this format and we see an opportunity for its adoption by other recommender systems.
This work was supported by grants VG1/0508/09, KEGA 028025STU-4/2010 and the Foundation of Tatrabanka. It is a partial result of the Research & Development Operational Program for the project SMART II, ITMS 25240120029, co-funded by ERDF. We wish to thank members of PeWe group, pewe.fiit.stuba.sk for valuable discussions and their help in experimental evaluation.