Popularity Bias in Online Dating Platforms : Theory and Empirical Evidence

Problem Definition: Motivated by the empirical evidence from a major online dating platform, this paper studies the existence of popularity bias in an online dating platform’s recommendations and its consequences for users’ likelihood of finding dating partners. Although the platform’s algorithm recommends popular users more frequently, popular users tend to accept messages from other users less frequently, resulting in suboptimal matching outcomes. Academic/Practical Relevance: While online dating has seen rapid growth in recent years, it has received relatively little attention from operations researchers. Our study contributes to the academic literature by studying the role of a two-sided matching platform on bias and fairness. It also contributes to practice by developing an accurate prediction model that can estimate users’ future decisions and by generating insights about how various user attributes affect users’ decisions to interact with others.Methodology: For our theoretical analysis, we model the online dating platform as a three-stage matching game. For our prediction analysis, we build a machine learning model that leverages recent advances in tensor recovery and interpretable tree-based models.Results: We find that the recommendations that maximize the platform’s revenue and those that maximize the number of successful matches between users are not necessarily at odds, even though the former leads to a higher bias against unpopular users. Unbiased recommendations result in significantly lower revenue for the platform and in fewer matches when users’ implicit cost of evaluating incoming messages is low. Popular users help the platform generate more revenue and a higher number of successful matches as long as these popular users do not become “out of reach.”Managerial Implications: Online dating platforms can use our theoretical results to understand user behavior and our predictive model to improve their recommendation systems (e.g., by selecting a set of users leading to the highest probabilities of matching or other revenue-generating interactions)