PROPOSED USER MODEL

The user's shopping behavior is classified as comparative shopping, planned shopping, or browsing-based shopping (Lu, 1999). In each of these behaviors, the needs of the user are different. During comparative shopping, the user's task is the selection of one item out of many. The user model can assist the user in the selection process. Filtering the results of the query posed to the database is a task that could be assisted. We do this filtering based on selected features. For example, suppose the user is looking for a motorcycle, and we know from his profile that he prefers Honda products. Then, we can eliminate all other motorcycles from other manufacturers. We can also disambiguate a query by filling in missing fields with values inferred from the user model.

In the case of planned shopping, we do another kind of information filtering. The filtering here is at a higher level of abstraction and more flexible because the user might have a "rough idea" about what he wants to buy in the future (e.g., a bicycle or a roller blade as next summer's sports hobby). In this case, we want to filter among a hierarchy such as departments, shelves and categories, before narrowing down to specific products.

Browsing-based shopping denotes shopping with a casual objective (window-shopping). Here, the main interaction problem is the limited size of the screen, since the user is looking for a panoramic view. Also we want to be able to identify the user's needs in this case and give interesting, personalized suggestions. In all the cases, the language, the content, and the presentation of the information should be adapted to the user's specific needs.

The model proposed by Abi-Aad (2001) contains three main types of information about the user:

The categories and subcategories of products the user is interested in. This knowledge can help filter information or personalize browsing on a general level. We refer to this as PIE (Preference Indication by Example).
The features of those products or categories. This knowledge can help compare items and predict the user's interests on a more specific level.
Any additional information about the user concerning these products, such as the reason for the user's interest in the product or his expertise in the domain. This knowledge about the user can help determine how to present the information. And, it can also help detect when an opportunity is interesting for the user. The user-centered additional information is also domain-dependent (e.g., expert in cars , professional skier); therefore, it is associated with products. We call it "user additional information".

A value denoting the degree of interest, ranging from “5 to 5, is associated with each of these types of information. In addition, a degree of certainty of each value is also included. Figure 1 shows the relationships among these three types of information as well as a part of the navigation graph. In the diagram, we have shown only the nodes with a value higher than a threshold. The PIEs form a natural hierarchy among themselves that is best represented as a DAG (Directed Acyclic Graph). It is not a tree because some products can belong to more than one category. This hierarchy helps build and refine the user model. It can help propagate information up-and-down the DAG and, thus, vary the relative notions of generic and specific. The information (values, features, and user additional info ) can be inherited from general categories to their subcategories. The information can also be overridden in specific categories to correct inaccuracies inherited from general categories.

Figure 1: A Fragment of a User Model

The second advantage of this hierarchy is being able to filter information on different levels, since the user's preferences might vary in precision. For example, consider the following PIEs hierarchy: "store ’ department ’ category ’ shelf ’ item". Each of these levels is, in fact, a subcategory of its parent category. When we are relatively sure of a "category" of product that the user would be interested in, then we compare values of PIEs at the following lower level in the hierarchy, the "shelf" level, in order to filter the information.

The features are represented by giving them values. So, the features are stored as attribute-value pairs (e.g., ˜color = red, ˜height = 5 inches, ˜wrinklefree = yes ). Also, features can be compared to values such as ˜height > 5 inches. Or, features are prohibited from having a certain specific value (e.g., ˜color red ). Features are associated with the products and, therefore, we associate a list of features with each PIE ” details about the car engine to someone who is an expert in cars - and only show external details, such as color , to a person who doesn't know much about car mechanisms.

During comparative shopping, item-wise comparison is usually made at a very low level, based on the features. During planned shopping, comparison is made at an intermediate level, based on the values of the PIEs. And, during browsing-based shopping, comparison based on the values of the PIEs is made at high and low levels, switching appropriately between them.