Personalized Recommendation Systems

Although research in personalized recommendation systems has become widespread only in recent years, the basic idea can be traced back to [64], discussed below with other work on conversational interfaces. [50] gives a more thorough review of recent research on the topic of adaptive interfaces and personalization. Several other adaptive interfaces attempt to collect user information unobtrusively. An interesting example is the CASPER project [61], an online recruitment service. The project investigates methods that translate click and read-time data into accurate relevancy information, given that the raw data is inherently noisy. Similarly, [35] describe a technique for learning web preferences by observing a user's browsing behavior. Another example is the ADAPTIVE ROUTE ADVISOR [65], which recommends driving routes to a specified destination. The system collects preferences about attributes such as number of turns and driving time on the basis of the user's selections and modifications of the system's proposed routes. While the ADAPTIVE PLACE ADVISOR uses constraint-based interaction to search for items, this is not the only interaction approach for item search. An alternative is taken by the candidate/critique, or tweaking, approach. Tweaking systems, such as the FIND ME suite [16], typically require the user to begin an interaction by filling in values for a few predetermined attributes. They then present an item, at which point the user has the opportunity to change some search parameters to try to find a more desirable item. [28] take a similar approach in their MATCHMAKING system. In addition, they exploit constraint satisfaction to manage the search. Neither the FIND ME suite nor the MATCHMAKING system, however, learns user models. A related system that does include a learning component is that of [71], which learns attribute preferences unobtrusively. While tweaking is a valid method, it is not appropriate, we feel, in an environment in which speech is the only interaction mode, since presenting the user with his options would be somewhat cumbersome. Even though our current system also presents options once the search is constrained, it limits the number of items presented. Even in a ``full speech'' version this does not seem onerous.