Personalized Shopping in the Web by Monitoring the Customer
Tanja Jörding & Stefan Michel
Multimedia Technology
Institute for Softwaretechnology II
Dresden University of Technology
{
tj4, sm5}@inf.tu-dresden.de
At present, electronic shopping in the Web is not as entertaining as it could be. Some customers do not have the preconditions for using multimedia presentations and avoid documents with audio data, other customers find textual descriptions boring and prefer product descriptions with video data or virtual reality (VR) worlds. Sometimes, customers have time to wait for some images; in another situation they want to get precise information as fast as possible. Because static product presentations cannot satisfy such different needs, we introduce TELLIM (inTELLIgent Multimedia), a system that generates individual documents at runtime such that every customer gets a presentation that appeals to him and that meets his actual interests.
Figure 1: Example for individual product presentations.
Figure 1 shows in the middle image how a car is presented at the beginning of a session. Depending on the user behaviour, the next presentation is generated. For example, if the user has chosen the text in the first presentation and has not enlarged the image or used the VR-world, the next presentation would be generated as in the left image. If the customer has not chosen the textual links, but navigated in the virtual world and enlarged the image, the presentation of the next car would be like the right image of Figure 1.
The adaptation process in TELLIM is divided into three parts [JM98]. In the first one, the system monitors interactions on the client-side. These are interactions with single presentation elements, e.g. what links he chooses, if he starts audio or video players or interrupts the downloading of images. The system evaluates such behaviour and decides if the customer was interested in the different presentation elements. Based on these evaluated elements, in a second part, an incremental algorithm learns the preferences of the customer considering three kinds of attributes: kind of medium, downloading time and a set of content attributes. Depending on the prediction of the learning algorithm, a last component generates the multimedia presentation at runtime. Therefore, the system integrates the supplements which the customer likes to use in the presentation and adds other elements only as links. This means that every piece of information concerning the product is visible for the customer, but preferred elements are presented in the foreground and uninteresting elements of the product description are presented only as links in the background.
A common element in many Internet data transfer protocols is an exchange of information which modifies the form and method of the data being transferred; in HTTP this is known as 'content negotiation'. Media feature descriptions and negotiation mechanisms are used in order to identify and reconcile the kind of information and the capabilities of the parties involved, e.g. higher or lower resolution images, the colour depth of the screen, paper or display size [Con98]. This approach concentrates on technical preconditions, but does not take preferences into account that depend on the contents of the presentation elements and on the actual situation of the customer.
Existing approaches for adaptive web interfaces use mainly an explicit knowledge acquisition by asking questions [Bla98], [Fir98]. However, in this application area, direct questions are problematic because customers are often distrustful and are not motivated to answer questions. Other systems try to monitor the behaviour of the user [Bro98], [Cha98]. But, this observation takes place only on the server-side and therefore does not provide enough information for an adequate adaptation. In TELLIM, the main part of the knowledge acquisition process happens on the client-side. Therefore, it is possible to monitor the interactions of the user with single presentation elements and to provide thus a basis for an immediate adaptation of multimedia product presentations.
For the adaptation process, there are mainly used rule-based systems where the adaptivity determinants are often hard-coded into the system [Bro98], [Bru94]. This approach is inflexible, as the system cannot easily customise to changing requirements of the individual user. Other web interfaces use collaborative filtering techniques [AKK97], [NP98], that use knowledge and preferences of a group for the individual. But, with this method it is problematical to initialize the system and to adapt the system for new products. Other systems use clustering algorithms [Cha98] to discover patterns of user behaviour. But, this operates only if the customer navigates long enough in the catalogue. An immediate adaptation is not possible. Therefore, the TELLIM system consists of an incremental learning algorithm, that change its rules shortly after single interactions to realize an immediate adaptation to actual preferences of the individual user.
Multimedia product presentations can consist of different media elements, like text, images, video, audio and VR-worlds. To understand the preferences of the customer, the TELLIM system offers different interaction possibilities, that may show the interest or disapproval of the customer without annoying him [Mic98].
Attached "Save" and "Print" buttons give the customer the possibility of capturing interesting articles. Stretchtext provides interested customers with access to more information. Extensive texts can be integrated with a scrollbar. These additional interaction possibilities can be implemented with the combination of Style Sheets and scripting languages – often referred to as "Dynamic HTML".
Similar mechanisms can be applied to images. Again, "Print" and "Save" buttons are added. Small images also have a button for maximizing (Figure 2). To give negative feedback, the customer has the capability to interrupt the downloading process.
Figure 2: Integrated Images
Due to the variety of audio and video formats, no uniform access to all products can be supported. Our solution concentrates on the file-based systems LiveAudio and the RealSystem 5.0 [Rea98] as an approach for streaming audio and video. The controls of the plug-ins can be embedded in HTML and monitored with the help of their Java interface, such that the system notices if the customer uses the "start" or "stop" button.
Virtual reality worlds are realized with the Virtual Reality Modeling Language (VRML) [VRM97]. To monitor user interaction in VRML-worlds it is necessary to access objects of a VRML-world. The VRML External Authoring Interface (EAI) [Mar98] allows this access from a Java applet in a web page if the VRML browser window is embedded on the same page. Using the EAI, a world-spanning "ProximitySensor" can be added, which notifies movements of the visitor.
In addition to the listed interactions, the system computes if the time spent on an presentation element was more than a certain threshold, to get another hint that the element met the preferences of the customer. With a set of rules, the system then evaluates for every element if the customer was interested or not. Examples of these rules can be seen in the following table.
|
Kind of medium |
Kind of integration |
Minimal time exceeded |
Interactions |
Evaluation |
|
text |
integrated |
yes |
scrolled |
positive |
|
image |
via link |
maximized, printed |
positive |
|
|
image |
integrated |
interrupted downloading process |
negative |
|
|
video |
integrated |
no |
no |
negative |
|
audio |
via link |
yes |
play and stop |
positive |
|
... |
... |
... |
... |
... |
In a next step, an incremental algorithm based on CDL4 [She96] uses these evaluated presentation elements as examples for learning the preferences of the customer. Therefore, the algorithm considers the kind of medium, the content attributes, and the downloading time as attributes for the presentation elements. The result of the algorithm is a decision list that can be interpreted as a list of rules, e.g.
then the customer may be interested.
and the kind of medium is not "video", then the customer may not be interested.
The chosen algorithm works incrementally, i.e. when the customer leaves one product presentation to go to another product, the algorithm gets the interaction data of the customer and updates its rule base. If there are contradictory data, the algorithm prefers the actual preferences of the user. These properties are important because many customers navigate only for a short time in the catalogue. To satisfy their needs, the adaptation of the presentation should happen quickly. Additionally, while navigating in the product catalogue, the preferences of the customers can change. Therefore, it is desirable that the system reacts immediately.
Depending on such learned rules and with the help of dynamic templates, the new documents are generated at runtime. Figure 3 shows the architecture of the TELLIM system, especially of the observer component on the client-side.
Figure 3: The architecture of the observer in the TELLIM system
The presentation elements are integrated in a dynamic HTML (DHTML) page. They are implemented in different ways. Images and texts are realized as DHTML objects. Interactions with the customer can be monitored with eventhandlers of JavaScript. Audio, video and VR-worlds are integrated with Plug-Ins. These Plug-Ins have Java interfaces that facilitate the monitoring of customer interactions. All interactions are registered in the EventMonitor, a component that is realized as a JavaApplet. This component collects the observed interactions and then evaluates the presentation elements. Using Remote Method Invocation (RMI) [RMI98], these evaluations are written in a central history list which is stored in the database and which provides the example data for the learning algorithm.
In this paper we introduced TELLIM, a system that generates individual multimedia product presentations at runtime. We focussed on the knowledge acquisition process and described how the system monitors the interactions of the customer on the client-side and how this information is used for an immediate adaptation.
In a next step, we want to evaluate the implemented prototype. Therefore, we first want to test the system with generated data to simulate different user behaviour considering the influence of the following parameters.
Secondly, we want to test the system with real users. Therefore, we will prepare one normal product catalogue and one adaptive catalogue with the TELLIM system and compare the behaviour of the customers. We want to measure the following objective data that indicate the acceptance of the customer:
Additionally, we will make questionnaires to find out the subjective opinions of the customers concerning the attractiveness of the presentation, the convenience of navigation, the downloading time, the amount of information, and the fulfillment of their expectations.
[AKK 97] Alspector, J., Koicz, A., Karunanithi, N. (1997), Feature-based and Cli-
que-based User Models for Movie Selection: A Comparative Study, in: User Modeling and User-Adapted Interaction, 1997, 7 (4), pp. 279-304.
[Bla98] B.H. Blackwell Limited (1998), http://bookshop.blackwells.com/
[Bro98] Broadvision, Inc. (1998), http://www.broadvision.com/
[Bru94] Brusolovsky, P. (1994), Adaptive Hypermedia: An Attempt to Analyze and Generalize, in: P. Brusilovsky, P. Kommers & N. Streitz (eds.) Multimedia, Hypermedia and Virtual Reality: Models, Systems, and Applications, Springer.
[Cha98] Charles River Analytics Inc. (1998) Open Sesame, http://www.opensesame.com
[Con98] Content Negotiation (conneg) Working Group (1998), (work in progress), http://www.imc.org/ietf-medfree/
[Fir98] Firefly Networks, Inc. http://www.firefly.net/
[JM98] Joerding, T. & Meissner, K. (1998), Intelligent Multimedia Presentations in the Web: Fun without Annoyance, in: Proccedings of the Seventh International World Wide Web Conference (WWW7), Brisbane, Australia, long version as internal report, http://www.inf.tu-dresden.de/~tj4/reports/tellim1.html.
[Mar98] Marrin, C. (1998) , Proposal for a VRML 2.0 Informative Annex - External Authoring Interface Reference, Silicon Graphics, Inc., (letzte Änderung 7.7.1998),
http://www.vrml.org/WorkingGroups/vrml-eai/proposals/chris/ExternalInterface.html[Mic98] Michel, S. (1998), Implizite Wissensakquisition für die Realisierung von adaptiven multi medialen Präsentationen im WWW, Master Thesis at the Dresden University of Technology.
[NP98] Net Perceptions, Inc. (1998), Building Customer Loyalty and Profitable 1-to-1 Customer Relationship with Net Perception’s GroupLensTM Recommendation Engine. http://www.netperceptions.com/product/white-papers.html
[She96] Shen, W.M. (1996), An Efficient Algorithm for Incremental Learning of Decision Lists. Technical Report, USC-ISI-96-012, Information Sciences Institute, University of Southern California, http://www.isi.edu/~shen/papers_by_date.html
[Rea98] RealNetworks Inc. (1998), Get Real G2 beta for RealAudio and RealVideo now, http://www.real.com/
[RMI97] Sun Microsystems, Inc. (1997), Remote Method Invocation Specification,
http://www.javasoft.com/products/jdk/1.1/docs/guide/rmi/spec/rmiTOC.doc.html.[VRM97] VRML Consortium, Inc.(1997), The Virtual Reality Modeling Language - International Standard ISO/IEC 14772-1:1997,
http://www.vrml.org/Specifications/VRML97/, letzte Änderung 10.1.1998.