Traditionally, user modeling in adaptive hypermedia systems is based upon monitoring requests of resources on the server. This information, however, is relatively vague and may lead to ambiguous or uncertain conclusions. One approach to improve granularity and accuracy in user models is monitoring \client-side" user interactions in the browser. Despite their steadily growing importance, especially along with the emergence of Web 2.0 paradigms, up to now such interactions are hardly being monitored. This paper shows how, by means of Web 2.0 technologies, additional information on user interactions can be retrieved, and discusses the bene ts of this approach for user models (like accuracy, granularity and immediate availability of data out of the request-response cycle) as well as for subsequent adaptations.
Traditional user modeling techniques of adaptive hypermedia systems (AHS) monitor requests of resources on the server to make assumptions on users. However, the fact that a document has been requested might be an insu cient source of information to determine whether it has been read. After servers have transmitted a hypermedia document, there is usually no further feedback to tell whether it is still open or which of its sections is being read. If a browser's request to the server (which refers to pages as a whole) is the only source of information available, then one cannot make assumptions in relation to the document's comprising parts or sections. Therefore, documents have to be treated as atomic items.
Additional information may be retrieved by calculating the time between
requests (e.g. in Knowledge Sea [
In contrast to these approaches, the one put forward in this paper aims to improve the monitoring process itself, and suggests a way to asynchronously monitor client-side user interactions, which user models and consequently AHS should bene t from. Although, especially with the emergence of Web 2.0, the importance of asynchronous client-server communication and client-side interactions have been increasing, they are hardly being used in terms of user modeling at the moment. 2
Instead of exclusively monitoring requests on server-side, interactions could as well be monitored directly within the browser. 2.1
As long as observation is reduced to server-side monitoring, user interactions
that do not cause browser requests can not be observed, which limits possible
assumptions on user behavior. Some of the questions that can not be answered
due to the lack of information on client-side interactions are [
As client-side monitoring provides more information on the users' interactions, it
allows additional assumptions and increases the level of certainty for conclusions.
This might help to improve both models and resulting adaptations.
Improving Accuracy of Existing User Models One evident example where user
models can be improved is the assumption of a user \having read a page".
Traditional systems make use of the information that a document has been requested.
Some more advanced systems consider the time of a subsequent request to
identify the \time spent reading" [
Independence of the Request-Response Cycle If the transmission of data depends
on subsequent requests, information is lost, if users leave a page. When using
browser history or parallel windows for navigation, incorrect data is transmitted.
Asynchronous monitoring is able to overcome these restrictions by retrieving
data out of the request response cycle, and additionally results in a model always
being up-to-date. Moreover, if changes in the user model require immediate
adaptations, this can be achieved as well. As the communication takes place
asynchronously, the system might o er help, highlight important elements or
provide any other (non-distracting) adaptation while the user is still interacting
with the page [
Attempts concerning client-side monitoring have already been made by Goecks
and Shavlik [
Due to technical boundaries (workarounds like refreshing iframes, Java
applets or ash movies have hardly been used) both monitoring and adaptations
were traditionally bound to a strict request-response cycle. In order to overcome
these limitations, custom browsers have been developed (like \AVANTI" [
Within \The Curious Browser" [
With the emergence of Web 2.0 and asynchronous web technologies like
AJaX (running in the background, not requiring additional plugins) some
limitations for continuous and unobtrusive monitoring of user interactions are now
addressed. Using these technologies not only for presentation, but also for user
monitoring has already been suggested for help systems by Putzinger [
As Web 2.0 technologies extend the range of what is technically possible, it
is now up to researchers to answer questions that could not be addressed earlier.
This includes increasing granularity and accuracy of user models [
In order to monitor user interactions inside the browser without requiring additional software, client-side scripting languages are required. Although there are several possibilities, for the system currently being developed JavaScript has been selected as the most appropriate choice, as it is widely supported and used. AJaX is used to send the collected data to the server as it is unobtrusive, not bound to the request-response cycle and natively supported by modern browsers. 4.1
JavaScript already provides a set of events that can easily be monitored. The current system uses various mouse and keyboard events, scrolling, window resizing, window blur, window focus and document ready.
However, some of these events occur too frequently for continuous monitoring, which is especially true for mouse moves and scrolling events. They are an indicator for interest and need to be monitored. For a single mouse move the browser may generate a large amount of JavaScript events, which would cause a lot of network tra c if every event would be sent to the server. Moreover, the area of the mouse move is of interest and not the exact route and all mouse positions. Therefore, these events need to be preprocessed and aggregated (e.g. all mouse-move or scrolling events within a small period of time are combined to a single one). In order to achieve this, the library developed along with the approach put forward in this paper supports the de nition of custom events. They can also be used for decreasing the level of detail (e.g. if exact positions are not required), adding supplemental information (e.g. based on the event context) and for creating specialized monitors. Additionally, custom events are necessary, because not everything that is monitored on the client side directly causes a JavaScript event. In the case of selecting text (which needs to be monitored e.g. to identify text tracing), di erent events and document states have to be monitored to be able to tell, when a \text selected" event occurred. Combining this event with additional mouse or keyboard events can show if a piece of text has been copied for further processing. Furthermore, custom events may use their own triggers, e.g. timers for events of a temporal basis.
The data provided by JavaScript events may be both too detailed and/or insu cient. In order to determine parts of a page a user has been interacting with, exact mouse positions hardly serve, if the content is not bound to xed positions or dimensions. Di erent font- or window sizes may change the absolute positions of content. In order to overcome these limitations the concept of page fragments is being introduced. 4.2
In order to be able to not only use a general \level of activity" within a page, but to treat parts of a page di erently, events need to be mapped to semantic meta information, and/or their spatial location needs to be identi ed.
One approach is to analyze the text a user is currently interacting with. For example the event triggered on text selection contains information on the currently selected text, which may be used for keyword analysis. For other types of events, like mouse-over events, analyzing keywords may be less e ective. These events are better suited to indicate activity within an area, than to assume the user is particularly interested in the words being hovered above. Furthermore, mapping events to keywords does not di erentiate between parts of a page using the same keywords and makes it therefore di cult to tell what percentage of a page a user has read. Despite being useful for some types of events, exclusive use of this approach is not able to address some of the di culties mentioned in previous sections.
Monitoring all HTML DOM elements and mapping events to them brings about problems as well. If a hypermedia document is highly structured in terms of markup (an element may contain a single letter), the result might be too detailed and consequently not useful. For some interaction types the exact position is only an approximation for the locus of attention. A user does not need to interact with every single element while reading it. Therefore, monitored areas must not be too small or detailed. On the other hand, within unstructured documents a single HTML element may contain the whole content, which results in treating the page as a single unit again.
Splitting hypermedia documents manually and providing semantic meta data for each section might be e ective, but requires additional authoring e ort and might therefore cause acceptance problems. The solution to be chosen should o er the possibility of manual sectioning, but provide its advantages also for unstructured contents.
The suggested solution tries to meet these demands by introducing \page fragments". By default, a page is vertically split into a number of k sections, each representing 1=k of the height of the document. Changing the width of a browser window may change the height of the document, but the height of the fragments will change as well. For k = 5 the third fragment will represent the center of the page; everything between 40% and 60% of the total height of the page. As this percentage will not change through resizing or using di erent font sizes (although the absolute postitions might change), this approach might be better suited for locating events. For some items, like images or spatial layout elements, resizing in uences exact positioning more strongly; but nevertheless, the approximation should be su cient.
One bene t of this approach is the independence of the structure of the content. Even unstructured documents may be segmented this way. For structured contents Therefore, no additional adaptive authoring e ort is required.
The actual size of a fragment depends on the height of the document and on the value chosen for k. The higher the value for k, the higher the level of detail. However, if k is too high and/or there is little content on a page the height of one fragment might be less than the line height, which means that clicks on the top of a word are treated di erently from clicks on the bottom of a word. Determination of optimal values of k for di erent circumstances is part of the ongoing work. As each fragment represents a xed percentage of the page and each fragment is monitored separately, it becomes easier to tell how much of a page has been read; based on how many fragments are regarded as \having been read". Although the claim something has been read still remains an assumption, having gone through the whole page, spending a reasonable amount of time on all of its parts and interacting with them should be a stronger indicator for reading than simply requesting a document and possibly spending time on it.
Although using fragments instead of absolute positions to retrieve spatial information results in a loss of precision, this approximation should be su cient for most applications, as e.g. mouse positions themselves are only an approximation for the locus of attention.
Within relatively static documents a fragment represents not only some percentage of the page, but also the actual content, which is important to determine interests. However, if parts of a page are adaptively included, the text within a fragment might shift, which requires a di erent technique to identify the actual content a user has been interacting with. Moreover, if semantic information is already available, it would be interesting to use it. Therefore, in addition to the prede ned fragments, \custom fragments" can be de ned, i.e. any HTML element can be speci ed as an additional fragment. Such custom fragments may be images (if they should be treated di erently), sections automatically detected by means of topographical information (e.g. using headlines as separators) or elements that are already mapped to concepts or semantic meta information. When assembling hypermedia documents from multiple sources (e.g. due to conditional fragment inclusion), fragments can, during assembly, automatically be denoted and semantically characterized on the basis of their source, their \function" in the assembly, etc..
Generally, custom fragments are treated like prede ned ones, so events are mapped to them as well (they can be mapped to more than one fragment). For instance, the time spent on images can be automatically calculated without requiring further modi cations. Moreover, events can be de ned for single fragments only, e.g. to specify the required reading time for each image separately (based on the actual content) and to trigger an event, if the time is exceeded and the image is regarded as visited. 5
Having monitored the user's interactions, the next step is to use this information within an AHS. 5.1
The acquired data can be used directly within existing AHS, for instance, by replacing values in the user model. As an example, the values for \having read a page" or \knowing a concept" may be rede ned within the user model by something more accurate, like \having spent at least x seconds on each page fragment". In this case, existing adaptive content might be used as is, while still pro ting from more adequate data. Secondly, the level of certainty for attributes within existing user models may be increased. A third example on how to make use of the available information (using existing adaptation engines) is the de nition of new and more granular adaptation rules like \probability of having read at least half of the page is greater than 80%". This helps authors of adaptive content to use the same authoring process as before, while still bene ting from user modeling based on client-side interaction monitoring. 5.2
\Social navigation" has been de ned as \moving `towards' a cluster of other
people, or selecting objects because others have been examining them"[
Using the activity information on single page fragments, social navigation
support can be provided not only on the level of documents and links, but also to
highlight sections of a document. Fragments with a high number of interactions
and reading time might be the most attractive and relevant ones. As users tend
to follow highlighted links [
Moreover, social navigation support may help authors to improve their contents by noticing, which parts of a document users found interesting. If within a block of important content users tended to spend time on the rst lines only and skip the rest, the content might be too di cult, already known or not interesting enough. Important sections that tended to be skipped might need to be improved. 5.3
One example where information on custom fragments opens new perspectives is the identi cation of learning styles. If images are treated as special fragments the system may identify the time spent on pictures and the interactions performed on them. These values may be compared to the average number of interactions / time spent on other parts of the page for the current user, which should allow for conclusions on the user's preferred learning style.
Having included this information into the user model, course material may be adapted accordingly. Contrary to traditional systems asking for learning style preferences, this approach might help to identify the learning style automatically by monitoring user interactions. 5.4
Perkowitz and Etzioni [
A JavaScript library for asynchronous client-side user monitoring as described in
section 4 has been developed. The implementation aims to be both easy to use
and simple to extend with respect to client-side observation as well as server-side
data processing. Existing courses can easily be provided with the core
functionality of the library, while still allowing experts to add advanced features. Ongoing
work on prototypes aims to provide and improve sample implementations
according to the proposals in section 5. In order to provide the results of this work
for a larger audience, the system is currently being integrated into a version of
AHA! [
Evaluation work planned for the immediate future aims to determine an appropriate value k for the number of fragments a page will be divided into. It is presumed that in order to have comparable data, k should be constant within a system. However, for di erent types of systems a di erent k might t best, depending on the actual content and its length. For example, in an AHS presenting slides (generally tting in one screen), a lower k might be su cient, in contrast to one providing e-books, where a page may contain a whole section or chapter. Two approaches will be used to determine the optimal k within speci c contexts. The rst one will start with a high amount of fragments. Values for neighboring fragments will be compared in order to nd out, how many signi cantly di erent sections this number could be reduced to. A second approach would allow users to manually adjust the value for k, so that it subjectively results in the most appropriate user model.
Another challenge is to determine suitable weights for interaction types when combining di erent types of interactions to a \level of activity" for speci c page fragments. A model capable of theoretically characterizing but also quantifying di erent types of activities will need to be selected and applied. The quanti ed \level of activity" resulting from said model will have to be veri ed with real end users' subjective perceptions of activity.
At a more general level, to evaluate the overall approach put forward in this paper, one would have to ascertain that: (a) it results in a more ne-grained and signi cantly more accurate user model and, (b) that this higher level of detail and the improved accuracy result in better adaptations.
For the tests, a version of AHA! based on traditional modeling techniques will be compared with an extended version analyzing client-side user interactions. To check the accuracy, a summative evaluation will be done through user studies. Students will work on their AHA! courses and will have to evaluate the two user models and decide which one they nd more appropriate. Care will be taken to ensure that users do not just select the model presenting the values most favorable to their person (e.g. showing higher values for knowledge for a large number of concepts). Alternatively, only one model could be used, presented and evaluated and the average evaluation of both models would then be compared. Another scenario would be a self-evaluation of students and a comparison of these results to both models.
The work reported in this paper is funded by the \Adaptive Support for Collaborative E-Learning" (ASCOLLA) project, supported by the Austrian Science Fund (FWF; project number P20260-N15).