Effective and diverse POI recommendations through complementary diversification models

Abstract

Nowadays, recommender systems play an important role in several Location-Based Social Networks (LBSNs). The current advances have considered the trade-off between accuracy and diversity to help users to discover and explore new points-of-interest (POI). However, differently from traditional recommendation scenarios, other equally relevant dimensions (e.g., social and geographical user information) have to be considered to understand how the characteristics of services offered by each POI fit the user needs. Specifically, this work sheds light upon naive failures introduced by traditional recommendation methods while they handle this trade-off between diversity and accuracy in POI recommendations. We hypothesize that some efforts on POI recommendations somehow are deviating from basic learnings from the area. In this context, this work addresses four characteristics inherent to the POI domain that previous efforts have failed to recognize: (1) POI categories and locations are complementary dimensions of diversification that should be simultaneously addressed; (2) Diversity is a complex concept that should be modeled by distinct and non-orthogonal models; (3) Distinct users have different biases and willingness to move to fulfill their needs; (4) POI recommendation is a multi-objective task. In order to demonstrate the gains of properly addressing these aspects, we also propose DisCovER, a straightforward re-ordering method that linearly combines geographical and categorical diversification. DisCovER results demonstrate that even simple strategies to exploit simultaneously these complementary dimensions can increase diversification while keeping accuracy high. Differently from state-of-the-art diversification methods, DisCovER does not penalize any quality dimension in favor of others. It allows us to discuss future directions towards more robust user modeling and preference elicitation in POI domains.