Exploration Principles for Decision-Making Systems with Binary Feedbacks

The challenge for a recommender system is to explore a complete solution space, preventing informational ’filter bubbles’ and ensuring a diverse range of recommendations. This paper introduces five exploration principles to better understand users and items while taking limited risks, as poor recommendation can also diminish user trust. The five principles apply to any type of recommendation system and are illustrated using a recommendation system for tourist points of interest (PoI). A novel interface was tested with open data and permitted to gather a preliminary set of user feedbacks. based on open data and a novel interface. Exploration raises two questions: which items should a given user be exposed to, and which users should be exposed to a given item. This latter question is studied with different exploration strategies, in order to check whether certain populations of users help quickly and efficiently learn about lesser-rated items. Synthetic data were used to compensate for the relative data scarcity of newly launched recommendation systems. Random user selection often proved sufficient, which can limit constraints when determining which items should a given user be exposed to, thus simplifying the exploration process. The principles outlined extend beyond PoI recommendation, having potential applications in industrial contexts such as simulation model recommendation and system parts suggestion.