Extended Abstract: A design for a tourist CF system

The use of computer supported travelling in the tourist industry has been steadily increasing and has recently attracted considerable interest. Tourism is in many ways the domain most closely connected with personal preferences and by definition connected to (physical) mobility. Hence, not surprisingly personalised location-based information systems are very suitable for this domain. The modern tourists do not only require general guidance and information but also information specifically tailored to their personal preferences. Local guides and guided tours cover many tourists’ needs by customising tours. Yet, a location-based personalised recommender systems offers a supplement to the available customised services. Recommender systems are designed to help users cope with vast amounts of information, and they do so by presenting only a certain subset of items that is believed to be relevant for the user. The typical tourist will not linger long in any location. Hence, a location-based information system will not be able to effectively learn the idiosyncrasies of any single tourist. This is a challenge when dealing with recommender systems, as they (most often) rely on a classification of the user and the information it is attempting to recommend. Not having sufficient information to give good recommendations to a new user is known as the cold-start-user problem. The cold-start-user problem can to some degree be alleviated by employing user models. However, building user models requires (sufficient) knowledge about the specific user. Acquiring this knowledge is subject to the knowledge bottleneck problem. That is, it is time consuming (for the user) and not necessarily easily accessible. A key question is therefore what type of information to query from a user, to what extent should information be collected, and how should the user information be exploited when the system gives recommendations. In this abstract we give the conclusions of a structured literature review [3] designed to answer these questions. The literature review focuses attention to CF models combining Bayesian networks with user modelling as a means of mitigating both the cold-start-user and knowledge bottleneck problem.