Abstract
Recommender Systems are popular tools that automatically compute personalised suggestions for items that are predicted to be interesting and useful to a user [24, 17]. For instance, in the music domain recommender systems support information search and discovery tasks by helping the user to find music tracks or artists that the user may not even know, but he will like [7, 15, 14].
Recommender systems accomplish their functionality by explicitly requesting users to enter their preferences and by tracking users' actions and behaviours, which implicitly signal users' preferences. Then, they aggregate these observation data and build predictive models of the users' future interests. Several techniques have been proposed to model user preferences and generate recommendations for them. But, ultimately, most of the implemented systems use content-, collaborative- or social-based approaches, or even more often, hybrid combinations of these three basic approaches [6].
In addition to long-term interests, which are normally acquired and modelled in RSs, other session specific factors do influence the user's response to the suggested items and therefore should be taken into consideration. These factors include: the ephemeral needs of the users [21, 19], their decision biases [8, 25], the context of the search [10, 18] and the context of items' usage [1]. However, appropriately modeling the user's preferences and behaviour in the possible various and diverse situational contexts and reasoning upon them in order to identify useful, convincing, diverse and relevant recommendations is still challenging. Major technical and practical difficulties must yet to be solved.
First of all, one should parsimoniously narrow down the various types and the number of contextual dimensions that the system should model to those that actually influence the user decision making processes [2, 23]. Then, it is important to understand the dynamics of the impact of such contextual dimensions on the user preferences and the decision-making process [8]. This impact is strongly coupled with the full interaction design of the system [5, 16]. Moreover, it is important to implement technical solutions that enable the system to continuously acquire context-dependent user evaluations (e.g., ratings) for the suggested items, during the full life cycle of the system [20, 11, 12, 22]. Finally, one must embed the contextual dimensions and leverage the acquired data in a recommendation computational model [3, 9], while dealing with the typically very limited knowledge of the system for the users, the items and the contextual situations [4, 13].
These topics will be illustrated in the talk, making examples taken from the recommender systems that we have developed in the tourism and music domains.