IVIany current web-based information systems need to adapt to individual user characteristics in order to streamline common tasks and enable users to use them efficiently. In this paper we describe an approach to user characteristics acquisition based on automatic analysis of user behavior thus minimizing the amount of necessary user involvement. We stress the importance of appropriate data collection and propose a novel approach to logging, in which we preserve the semantics of logged evelrts. This has a significant inrpact on the successive log analysis, which focuses on the estimation of user characteristics from user navigation.
Present-day web-based information systems increasingly take advantage of personalization as vitai means of streamlining user experience, which is seriously hampered both due to the growing complexity of applications and the constantly increasing size of the available information space. The demand for generic adaptive systems results in the demand for generic methods and software tools, which perform automatic construction and maintenance of individual users' models. However, this is a compiex continuous process of acquisition of user data and its processing with the corresponding user model update.
Our primary goal was to devise methods for automatic gathering of user characteristics with minimal user involvement that can be successfully employed for personalized navigation, content recommendation and/or filtering in a particular dornain (such as job ofi'ers or scientific publications). Consequently, wc deviscd ntethods that first acquire data about user behavior by observirrg user actions and logging them while preserving their semantics. Next, they analyze this data in order to discover meaningful user characteristics and lastly perform the corresponding user model update to add the newly discovered knowledge to the already existing knowledge in the user model. We evaluate our research in two domains: the domain of job offers and the domain of scientific publications. 2
There are two main approaches to user activity logging - standard web server
logs and client-side logs created by special software tools deployed on a user,s
computer. Web server logs are commonly used as input for various data mining
techniques tn Web UsageMi,ni,ng[
Unfortunately, web server logs do not provide enough information due to the caching mechanisms of web browsers and basic principles of the HTTp protocol, which is a low-level stateless protocol without clearly defined semantics of performed actions (u.g., GET and POST do not supply adequate information). Furthermore, a web-based adaptiue system with dynamic page generation can produce two similar or even equal log entries which result in completely different system responsesdue to changes in the user and domain models.
To address some shortcomings of this approach and to capture information
about client-side only interactions (e.g., hovering on tooltips) client-side iogging
methods were proposed. lVIost are either pure ciient-side only syste-r [S] ot
standalone desktop applications communicating with a server [
Several approaches to user characteristic discovery have already been
proposed' One important method of user characteristics discovery is ihe analysis
of user navigation. In [
While existing server-side and client-side methods of data acquisition can be used. they are severeiy insufficient since they fail to effectively capture the semantics of individual events for further processing. The analysis of user navigation and irnplicit user feedback is a promising method of user characteristics discovery' However, most approaches focus only on the discovery of user ratings and do not explore the reasons behind them. Here the comparison of displayed and rated concepts by means of concept comparison methods might yield more detailed and/or accurate user characteristics. Similarly, authors t iit use user ratings to discover rules expressing user preferences on ordered values. 3 Method for lJser characteristics Discovery Figure 1 shows the architectural overview of a web-based system employing the proposed method as an integrated set of cooperating software .o*por,.rrts [t3].
Personalizedpresentation Presentation . . . . ' . : ' . : . ' . . . . . . . .
Personalization
layer User modeling layer
User characteristic
evaluation plicatlonlogic la Data la
User CharacteristicsAcquisitionfrom Logs with Semantics 105 We combine and enhance both client-side and server-side logging approaches and create a comprehensive 1og of user actions while also preserving the semantics of individual actions (Figure 1, bottom right). Although server-side monitoring is reliable, it might miss time-related data about user interaction. Thus we use client-side logging as an optional source of time-precise information, since we have no direct control over the execution of ciient-side logging mechanisms (no data can be gathered if the user disables monitoring on the client side). Client-side logging. Client-side logging captures events that occur entireiy on the client-side (in the web browser) during user interaction with elements on the displayed pages and also captures precise time-related data about individual actions which might be missed by server-side logging. The monitored events include Load, (Jnload, Cli,clc,Mouseouer and Mouseout.
For each captured event, we collect the following data: type of euent, time when it was fi.redand contert of the captured event, e.g. what link was followed. F\rrthermore, additional events invisible to the server-side might be captured suclr as Change, invoked when the content of a form control changes. The sequcncc of such events indicates the order in which a user fills thc form. Another example is the Scroll event, invoked when the user scrolls the content of a page. Server-side logging. Standard web server logs are not suitable for the estimation of individuai user characteristics, since they require complicated preprocessing 11]but also lack the semantics of performed actions. We propose a method of server-side event logging that allows for the logging of events with defined semantics by both server-side and client-side tools and for the integration of these events into continuous streams of events for a particular user and user session. Nloreover, since only individual presentation/interaction tools "understand,, the setuantics of events that they process we propose them to be responsibie for the iogging of their own events by means of a common server-side logging tool.
Consequently, server-side logging aggregates events from various sources and stores their semantics which mostly include references to concepts from a domain ontoiogy. To further separate log analysis from log creation we devised a colrrmon euent ontologg that defines the semantics of individual events and their attributes. F\rrthermore, to evaluate the reasons behind user actions we also log thc logical display state of the user interface at the time when an event took place. This allows us to analyze user decisions (i.e., occurred events) based on what the user saw in the web browser interface (i.e., what her reasons were).
If for instance the system was using a faceted browser for navigation in the infbrmation space of job offers, one event could be to display details about a spocific job offer, while another might be to show oniy job offers in New york. For both of these events, the meaning of the action would be logged (i.e., URI of tlie respective event) together with attributes such as the URI of New york. Lastly, the logical display state would be logged indicating amongst others which otlter job offers were dispiayed.when the user chose the details of a specific one. 3.2
Log Analysis The purpose of log analysis is to process the acquired user activity logs, to identify meaningful user characteristics and to update the user model to reflect the newly gained knowledge. We focus on two main aspects - on the eualuat,ion of u'ser nau'igation in the available information space and on the eualuation of i'mplici't feedback. We assume the use of a navigation model that allows the user to tnove freeiy on a site, so her behavior can be influenced by her characteristics. Individual steps of the log analysis method are depicted in Figure 2. Data preprocessing. The proposed logging methods produce ,,ready-to-use,' data that makes preprocessing much easier compared to the preprocessing required when using standard web server logs. However, some additional transformations can be applied such as the removal of successiveidentical actions (e.g., when the user repeats an action due to slow system responseJ.
Pattern detection. Wc pcrforrtr the rnapping of everrtsonto a set of predefilcd behavior patterns - sequencesof successiveevents, for which we devised a special representation using event semantics. Each sequencemust occur a given number of times (count-of-occurence) to satisfy a pattern and can be defined as continuous which requires that the events are uninterrupted (i.e., the rule cannot span multiple sessions).tr\rrthermore, sequencescan be nested and each sequence or User CharacteristicsAcquisitionfrom Logs with Semantics 101
Heuristics Fig. 2. Overview of user characteristics acquisition. Data from presentation tools and client-side logging (1) are stored in a database of user actions (2). After its preprocessing (3) we try to detect occurrences of predefined patterns (4) and opbionaliy store intermediate results (5). lve determine the user model update based on the heuristics associated with the detected pattern (6). If a pattern represents implicit user feedback, we evaluate it (7) and estimate user preferences by comparing the respective concepts (8). Before the update of the user model, we evaluate user behavior consistency (g) which influences the confidence of the reveared characteristics (10). everrt czrn hurve a contttxt rvhich defirres firrther restrictions (e.g., which event attribute is required to change in the next event and which attribute must remain the same for all events and subsequences of a specific sequence).
The pattern in the following example is satisfied when a user logs into a
system and selects at least one restriction in a faceted browser interface (u.g.,
selects Washington state). It allows the user to select more than one restriction
if their PropertyUri remains the same and restrictlonUri is different (which
means the user is refining the restriction in the current facet, e.g. selects Seattle).
This pattern focuses on "first click behavior", right after the user's iogin, since
the first click usually has higher relevance to the user,s goal [
<attribut e s >http : / / t iit. sk#propertyUri</attribut < / c o n t e x t > <context type="http :/ / fiit. sk#Diff erentThanprevious")
<attribut e>http : / / t iit. sk#restri ct ionURI</attribut e> </context> </event> </sequence> (/sequence> Inconsistent behavior detection. lVe use the sequentialpatterns mining data mining method, which operates on previous user sessions.We consider the identified sequential patterns to be patterns of user behavior. The current user session is compared against these patterns to find out whether the current user behavior is "consistent" with the user's previous behavior patterns.
If this is not the case, one of the following heuristics is used:
1. If the time difference between sessionsis too trigh. we assume that the user's
characteristics have changed and lower the confidence of the characteristics
stored in the model (or even reset the whole model and start over).
2. Else we assume that the user is randomly browsing a site without following
a defined goai or is acting on behalf of somebody else. We thus significantly
Iower the confidence of characteristics discovered in the current session.
Feedback evaluation. If a pattern of implicit feedback was detected [
During the comparison we examine the relationship of the compared
instances based on the used class taxonomy. We also traverse all of their attributes,
which might be either data type or object type attributes. When comparing data
type attributes, a simple comparison of strings does not give satisfactory results
as the semantic dissimilarities in compared texts need to be evaluated [
User CharacteristicsAcquisitionfrom Logs with Semantics 109
Each characteristic stored in the model has a level of conf,dence- how reliable we think the estimation is, and a level of. releuance of the characteristic to a defined goal. Heuristics define how to modify these two parameters with support for separate update strategies for relevance and confidence and also for the use of hard limits on the extent of applicable changes. Confidence updates are also affected by the results of behavior consistency detection. 4
We evaluated the proposed method using the domain and user models in the domain of job offers and in the domain of scientific publications. For this purpose we implemented software tools to perform the respective tasks.
- Cli,ck tool realizes client-side logging, It captures events fired by a web browser during user interaction with elements on displayed pages. Since standalone applications are inflexible and pose a privacy threat, we implemented Cli,ck using JavaScript technology. - Semanti.cLogtool implements server-side logging by means of a web service.
It allows both presentation tools and Cti.ckto log data about user interaction, which are then aggregated into one common 1ogper user sessionand stored in a relational database. - LogAnalyzer tool performs the estimation of user characteristics from the acquired user logs. It continuously evaluates incoming events and performs user model updates, which are used to personalize the content and presentation of the system by means of presentation tools such as faceted browser.
Our approach has several vital advantages over existing solutions. The main improvement to logging is that the semantics of user actions and of the system itself are acquired and preserved for further processing by user characteristics estimation tools. These include both the relevant events of the server-side presentation/interaction tools and the purely client-side events caused by the user.
Presently, most log analysis tools need to "understand" the URL addresses and the associated parameters of individual presentation tools in order to "comprehend" the semantics in logs leading to tightly coupled tools. The addition of another presentation tool thus results in a (major) update of the log analysis tools. Our server-side logging with semantics circumvents this tight coupling of presentation and log analysis tools by providing a common representation of logged events from all presentation tools by means of an event ontology. Furthermore, virtually no preprocessing is required compared to the extensive one required to extract user and sessiondata for standard web server logs.
A significant advantage of our log analysis approach is its universality. We carr define rnore or less complicated sequencesof events with semantics typical for a particular application domain. Our initial evaluation shows that individual patterns depend on the used presentation method and/or navigation model and can be successfully reused in multiple domains if the presentation method remains the same (".g., a faceted browser). One more advantage is that each estimated characteristic contains information both about its relevance and its reliability of estimation. Personalization tools can take this information into account when using the characteristics stored in the user model for successiveadaptation.
Our future work wili include more comprehensive evaluation and estimation of more complex characteristics and combinations of characteristics. Moreover, support for (semi)automatic pattern modification and/or generation might improve user characteristic estimation accuracy and efficiency.
Th"is work was part'ially supported by th,e Slouak Research and Deuelopment Agencg under the contract No. APVT-20-007104, the State progran'Ln7eof research and deuelopment under the contract No. 1025/04r and the Sci,enti,ficGrant Agency of Slouak Republi,c,grant No. VG1/3102/06.