Personalization

Personalized user adaptive systems obtain preferences from their interactions with users, keep summaries of these preferences in a user model, and utilize this model to generate customized information or behavior. The goal of this customization is to increase the quality and appropriateness of both the interaction and the result(s) generated for each user. The user models stored by personalized systems can represent stereotypical users [20,64] or individuals, they can be hand-crafted or learned (e.g., from questionnaires, ratings, or usage traces), and they can contain information about behavior such as previously selected items, preferences regarding item characteristics (such as location or price), or properties of the users themselves (such as age or occupation) [46,64]. Also, some systems store user models only for the duration of one interaction with a user [19,73], whereas others store them over the long term [65,12]. Our approach is to learn probabilistic, long-term, individual user models that contain information about preferences for items and item characteristics. We chose learned models due to the difficulty of devising stereotypes or reasonable initial models for each new domain encountered. We chose probabilistic models because of their flexibility: a single user can exhibit variable behavior and their preferences are relative rather than absolute. Long-term models are important to allow influence across multiple conversations. Also, as already noticed, different users have different preferences, so we chose individual models. Finally, preferences about items and item characteristics are needed to influence conversations and retrieval. Once the decision is made to learn models, another design decision relates to the method by which a system collects preferences for subsequent input to the learning algorithm(s). Here we can distinguish between two approaches. The direct feedback approach places the burden on the user by soliciting preference information directly. For example, a system might ask the user to complete a form that asks her to classify or weight her interests using a variety of categories or item characteristics. A recent study [56] showed that forcing the user to provide ratings for items (movies, in this case) that they choose, rather than those that the system chooses, can actually lead to better accuracy rates and better user loyalty. However, users can be irritated by the need to complete long questionnaires before they can even begin to enjoy a given service, and the study was not in the context of a dialogue system but involved a simpler interaction. Another, slightly less obtrusive, form of direct feedback encourages the user to provide feedback as she continues to use a particular service. The second approach to acquiring user models, and the one taken in the ADAPTIVE PLACE ADVISOR, is to infer user preferences unobtrusively, by examining normal online behavior [33,61]. We feel that unobtrusive collection of preferences is advantageous, as it requires less effort from the user. Also, users often cannot articulate their preferences clearly until they learn more about the domain. A possible disadvantage to unobtrusive approaches is that users may not trust or understand the system's actions when they change from one interaction to the next. This could be addressed by also letting the user view and modify the user model [45]. Systems typically take one of two approaches to preference determination. Content-based methods recommend items similar to ones that the user has liked in the past [67,58,49]. In contrast, collaborative methods select and recommend items that users similar to the current user have liked in previous interactions [24,12,47,70]. Because collaborative filtering bases recommendations on previous selections of other users, it is not suitable for new or one-off items or for users with uncommon preferences. The content-based approach, on the other hand, uses the item description itself for recommendation, and is therefore not prone to these problems. However, content-based techniques tend to prefer the attribute values that users have preferred in the past, though they do allow new combinations of values. We feel that the benefits of a content-based approach outweigh the disadvantages; we discuss methods for overcoming these disadvantages and for combining the two techniques in Section 6.3. Ultimately, personalization is about how one can utilize a learned user profile to search for, identify, and present relevant information to the right user in the right way at the right time. User models have been utilized in recommendation systems for content processing and selection (information filtering), navigation support in web browsers [58], and choice of modality and style of presentation and interaction [15]. The ADAPTIVE PLACE ADVISOR adapts its information filtering and interaction behavior, since these are most relevant for our application and since the majority of the interaction is through natural language.