A contextual approach to improve the user's experience in interactive recommendation systems

Recommendation Systems have concerned about the online environment of real-world scenarios where the system should continually learn and predict new recommendations. Current works have handled it as a Multi-Armed Bandit (MAB) problem by proposing parametric bandit models based on the main recommendation concepts to handle the exploitation and exploration dilemma. However, recent works identified a new problem about the way these models handle the user cold-start. Due to the lack of information about the user, these models have intrinsically delivered naive non-personalized recommendations in their first recommendations until the system learns more about the user. The first recommendations of these bandit models are equivalent to a random selection around the items (i.e., a pure-exploration approach) or a biased selection by the most popular items in the system (i.e., a pure-exploitation approach). Thus, to mitigate this problem, we propose a new contextual approach to initialize the bandit models. This context is made by the information available about the items: their popularity and entropy. The idea is to address both exploration and exploitation goals since the first recommendations by mixing entropic and popular items. Indeed, this approach maximizes the user's satisfaction in the long-term run. By a strong experimental evaluation, comparing our proposal with seven state-of-the-art methods in three real datasets, we demonstrate this context achieves statistically significant improvements by outperforming all baselines.