主题｜Topic：Biased Personalized Recommendation under Compliance Rules
时间｜Time：2022.06.10（周五）｜June. 10th (Friday), 2:00-4:30 PM
地点｜Venue：文澴楼521会议室｜Meeting Room 521，WENHUAN
主讲｜Speaker
徐子彬目前任香港城市大学深圳研究院副研究员，毕业于美国南加州大学马歇尔商学院，曾任上海交通大学营销系助理教授和副教授。他的研究成果发表在 Marketing Science, Management Science, Journal of Retailing 等期刊，主持国家自然科学基金面上和青年等项目。
研究领域｜Research Interests
企业决策的量化模型，包括大数据下的个性化定价策略与偏好信息匹配，消费者隐私与数据管控，互联网平台治理与信息设计等
摘要｜Abstract
通过从消费者的行为和个人数据中获取优势偏好信息，Netflix 和天猫等数字平台可以使用个人化推荐来促进产品匹配。然而，推荐者可能使用有偏见的算法来引导消费者接受一个不太相关但更有利可图的产品，而这将引起消费者的怀疑，导致消费者拒绝推荐。在这项研究中，我们用空谈博弈模型建立了优势信息下的产品推荐的微观基础。在模型里，消费者仅部分了解自己的匹配价值，而行为数据和个人数据与匹配价值相关。我们考查了三种对消费者数据的合规法则：推荐者可以使用 (1) 仅个人数据、(2) 仅行为数据或 (3) 结合两种数据，来进行个性化产品推荐。我们的结果表明，在均衡状态下，更多的消费者数据会带来更精准的偏好信息，但未必会导致更准确的产品推荐。当消费者留存率较低时，推荐者可能宁可放弃优势信息。这是因为策略性的不准确产品推荐可以减轻消费者对推荐者滥用优势信息的怀疑，从而提升消费者接受推荐的概率。此外，当推荐者仅使用个人数据，消费者福利可能比用行为数据时更高，这意味着允许平台使用个人数据而非行为数据进行算法歧视有可能更有利于消费者。
By learning superior consumer information from both behavioral and personal data, a platform such as Netflix and Tmall can use targeted recommendation to facilitate product matching. However, the recommender may have biased incentives to steer consumers into accepting a less relevant but more profitable product, which may induce consumer suspicion that leads to rejection. In this research, we build a micro-foundation of cheap-talk recommendation for consumers who are partially informed of their match value, which is correlated with both the behavioral data and personal data. Specifically, we examine three compliance rules, under which the recommender is permitted to personalize recommendations using (1) only the personal data, (2) only the behavioral data, or (3) both data. Our results show that in equilibrium, more consumer data, even creating superior information, may not lead to more accurate recommendations. In addition, the recommender may prefer no superior information when the consumer retention rate is low. This is because strategic inaccuracy can alleviate consumer suspicion of abusing superior information and thus increase the acceptance rate. Finally, consumer welfare may be higher when the recommender uses only the personal data than the behavioral data, which implies that demographic targeting can be more consumer beneficial than behavioral targeting.
