"Where Far Can Be Close": Finding Distant Neighbors In Recommender Systems

Location and its corollary, distance, are critical concepts in social computing. Recommender systems that incorporate location have generally assumed that the utility of locationawareness monotonically decreases as entities get farther apart. However, it is well known in geography that places that are distant “as the crow flies” can be more similar and connected than nearby places (e.g., by demographics, experiences, or socioeconomic). We adopt theory and statistical methods from geography to demonstrate that a more nuanced consideration of distance in which “far can be close” – that is, grouping users with their “distant neighbors” – moderately improves both traditional and location-aware recommender systems. We show that the distant neighbors approach leads to small improvements in predictive accuracy and recommender utility of an item-item recommender compared to a “nearby neighbors” approach as well as other baselines. We also highlight an increase in recommender utility for new users with the use of distant neighbors compared to other traditional approaches.