User Models: Customizing E-Commerce Websites to the Context of Use

Abstract

This paper deals with the problem of the growing complexity of the Internet. We focus our attention on e-commerce as a domain of application with which we can experiment. In this e-commerce context, we note the following:

• User interfaces play an important role in achieving user acceptance.

• Queries usually return more matches than the user can consult, or fewer matches than expected.

• The user is “flooded” by unwanted and sometimes unsolicited information (e.g. advertisement banners that pop-up or appear as part of the main window of the browser).

• The information is sometimes very badly organized, which makes it difficult to read and scan through.

• Finally, some of the cultural and ethical values of shopping in stores are missing when shopping through the Internet (trust, honesty, negotiation, policy, etc.)

One possible way to solve these problems is through personalization of the interaction using user models and inferring based on such models. We define and describe a user model that helps the user in four different ways:

• 1. Providing personalized services to a particular user. For example, by filtering out the information that does no correspond to the user’s center of interest.

• 2. Presenting information in a way suitable to the user’s needs. For example, presenting the information in an appropriate language.

• 3. Providing proactive feedback to assist the user. For example, a hint message that pops up when the user is taking too long to perform a task.

• 4. Disambiguating user’s search input based on his user model. For example, filling in missing fields by anticipation in a query form.

These issues are not new. Many other user models have in the past focused on these issues. Their limitation is that they are applicable in narrow domains. For example they helped the customer choose between two different books. The user model we propose in this paper is applicable to relatively a broader domain, like virtual malls or departmental stores. For example our user model can predict that the customer at a particular moment is more interested in buying a TV set than a bicycle.

In order to define our user model, we started by analyzing the user’s behavior and we classified it into three types, which are “comparative shopping”, “planned shopping” and “browsing-based shopping”. Comparative shopping is when the user has a specific need in mind. He knows what he wants, and he will spend a short time looking for it, comparing prices, brands and options. Planned shopping is when the user has a rough idea about his needs and he is not in a hurry to get it. So he will spend more time looking occasionally for what he wants, and narrowing-down the different possibilities. Browsing-based shopping is when the user does not have any specific need. He just wants to go shopping to gather information, spend time, socialize, etc.

The user model we propose stores the knowledge about the user in a hierarchy best represented as a directed acyclic graph. Each node in the graph represents a certain category of items sold. To each node is associated also some information like a value denoting the degree of interest of the user in this category, some options and features along with their values (e.g. Color = Red), and some information about the user like the reason for his interest or his expertise in this domain.

The hierarchy in this model helps build the user model by propagating the information up and down the graph (inheritance). And also it helps predict the next move of the user while browsing through the information. The features associated to the nodes, help filter the results of a query. The information about the user, help the system decide when to notify the user about a certain offer, and also it helps decide how to present the information (i.e. to which degree of details).

In order to test the user model, we classified the user’s initiative to interact with the system into three modes, and we applied the user model in these three modes of interaction. These modes are: querying, browsing and being informed (proactive notification). In each of these modes of interaction, the user model can be used differently to personalize the interaction. For example when the user is querying a database of products, we can filter the results of his query or disambiguate his query based on his model. When the user is browsing through the store, we can predict his moves, and we can personalize the display based on his preferences. And also we can notify the user of special offers, based on his urgent needs.

The three types of identified shopping behaviors are explored in details. The difference among them has to do with the lengths of the time-window available for the user to complete the shopping activity, and the detail to which the user needs are known to himself/herself and to the system. Our user model is motivated by our perceived need to broaden the coverage of the domain of products while dealing with a virtual or electronic shopping mall. Ongoing research in our team is now focusing on capturing the knowledge about the user based on the parameters suggested in this paper.