Who is recommended matters: an investigation from a relational perspective

Since the performance of the platform can be largely influenced by the relationships between users, it is important to consider the effect of recommender systems from a relational perspective, which is overlooked in the literature. Our study complements the literature by distinguishing recommended candidates as either familiar or unfamiliar based on relationship embeddedness. Using a recommender system's launch on a freight exchange platform as a natural experiment, we employ difference-in-differences estimations to quantify the effects of recommending diverse candidates on user's acceptance and overall transaction frequency. Results show that recommendations with familiar candidates are more likely to be accepted, but lead to a lower user's overall transaction frequency. On the contrary, recommendations with unfamiliar candidates result in a higher user's overall transaction frequency, suggesting that the positive effect of the recommender system is primarily driven by recommendations with unfamiliar candidates. We attribute these findings to informative signals of more matching opportunities conveyed by recommendations with unfamiliar candidates, which elevate user's evaluation of the platform and stimulate more transactions. Our work contributes to the literature by highlighting a relational perspective in the design of recommender systems and the importance of diversity rather than accuracy.