The SeniorGezond Recommender: Exploration Put into Practice

Recommender systems suggest objects to users navigating a web site. They observe the pages that a user visits and predict which other pages may be of interest. On the basis of these predictions recommenders select a number of pages that are suggested to the user. By far the most popular recommendation strategy is to select the pages of which the recommender believes they are the most interesting for the user. However, various simulation experiments have shown that this strategy can easily lead to tunnel vision: the recommender keeps recommending elements that are very similar to the pages that the user has visited and never discovers interests in other topics. In this paper, we describe a recommender system that does not always recommend the pages that seem most interesting but that also recommends pages that represent other parts of the space of pages. This helps to obtain usage data about parts of the space that the user has not yet visited. Moreover, the recommended pages are less obvious and thus more surprising to the user. The recommender system is compared online with a hand-made recommender of a real web site. Results show that the new recommender was used more frequently. However, the pages reached through the recommendations are read more shortly than the pages reached through the baseline recommendations. An explanation is that the more surprising recommendations are used most frequently by users with unspecific information needs.