The Daily Me Versus the Daily Others: How Do Recommendation Algorithms Change User Interests? Evidence from a Knowledge-Sharing Platform
利用知乎从内容过滤转向社交过滤算法的准实验，发现算法变化使社交连接增加15%，但问题订阅和回答贡献分别下降20%和23%，用户社交兴趣更集中而话题兴趣更分散。
Recommender systems on online platforms are often accused of polarizing user attention and consumption. The authors examine this phenomenon using a quasi-experiment conducted by Zhihu, the largest online knowledge-sharing platform (or Q&A community) in China. Zhihu originally used a content-based filtering algorithm, which recommends content to users on the basis of the topics to which each user has subscribed. After more than a year, Zhihu moved to a social filtering algorithm, which recommends content with which users’ social connections are already engaged. The authors find that this algorithm change increased the creation of social ties by approximately 15% but decreased question subscriptions by 20% and answer contributions by 23%. The authors show that users’ increased social interests mainly involved following popular users, leading to a greater concentration of social interests on the platform. However, users’ topical interests became less concentrated, as popular topics received significantly fewer subscribers than unpopular topics. The authors explain these findings by exploring the underlying mechanism. They show that compared with content-based filtering algorithms, social filtering algorithms are more likely to expose general users to content consumed by their followees, who are more interested in niche topics than general users are.