|M.Sc Student||Naidich Alexander|
|Subject||Exploiting Prior Knowledge in Reasoning about User|
|Department||Department of Industrial Engineering and Management||Supervisor||Professor Carmel Domshlak|
|Full Thesis text|
The product catalogs of online merchants and information providers grow continuously, and with them grow the number of lay users accessing these catalogs. While keyword search provides them with some means of access to the catalogs, user queries in the shopping context are typically more complex than in Web search. Using the purchase of a used car as an example, a sophisticated catalog search-aid system is envisioned in to allow users stating preference statements like ”I like ecologically friendly cars”, ”For a sport car, I prefer red color over black color”, ”This car is better than that car”, ”This car would be better in red”, etc. The system should use this qualitative preference information to create an effective user model, and utilize the created model to guide the user to the most relevant parts of the catalog.
A robust solution to this problem based on high-dimensional utility function compilation has been introduced in the work of C. Domshlak and T. Joachims "Unstructuring user preferences: Efficient nonparametric utility revelation". In our work we extend this approach to better deal with the (typical in practice) situations in which the user in question provides only a small number of preference statements. The basic technique is this of collaborative reasoning about the users’ statements, and the basic assumption is that the new user is likely to be somewhat similar to a group of other users.
First, we propose and evaluate a concrete extension of the approach of the work of C. Domshlak and T. Joachims to the setup of collaborative reasoning. Second, we extend further the theory to a more generic case when there could be a number of different sub-communities of users when each group has its own similarity among group’s members. We call this extension ”users clustering”. We discuss different semantic and computation problems that exist with this extension, like how to decide on number and types of clusters, and how to decide what ”type” of user we are dealing with. Finally we propose 2 possible heuristics that can be used when dealing with user clustering.
On the side of empirical evaluation, we show that the extended methodology provides a significant improvement over its basic counterpart.