A novel similarity measure for group recommender systems with optimal time complexity

Abstract

Once we subscribe to an e-commerce portal, or to a social media website, we interact with multiple brands and with content from numerous providers. However, a unique user profile is created, containing all our preferences. Suppose that a company wants to understand who are its customers. It wants to treat costumers as a target, and understand what campaigns the company should run on them. On the one hand, an approach that clusters the users and performs group recommendations would be useful, while on the other hand, a generic user profile would not be helpful, since the preferences in it are not specific for a brand. Hence, we have to determine multiple user clusterings (one for each brand). This task makes the problem of producing group recommendation challenging, since little and very sparse information about the users is available, and for each pair of users we have to detect as many similarities as the brands existing in the system. To tackle this problem, in this paper, we introduce a novel and optimal measure to compute the similarity between users, based on Kolmogorov complexity. Further, we test it in the group recommendation scenario. The results show that our similarity measure can provide similar accuracy when compared to classical measures, but with significant performance gains, having a strictly lower time complexity than the state-of-the-art similarity measure.