社交网络上的注意力拍卖
Auctioning Attention on Social Networks
浏览论文内容
中文总结 AI 辅助
针对社交媒体注意力分配冲突，提出基于拍卖的信息流构建机制，引入税收政策平衡生产者与消费者福利，在实证与合成网络模拟中提升生产者福利并优化注意力分配公平性。
中文摘要 AI 辅助
社交媒体推荐系统引发了多方冲突：内容生产者、内容消费者、平台运营者以及社会压力都试图将注意力分配导向对自身有利的方向，同时又面临着相互竞争的社会压力。生产者可能感受到优化推荐算法的压力，消费者则可能接触到带有极化、错误信息等负外部性的内容，平台则以最大化用户参与度为目标，这往往导致消费者过度消费内容。政策制定者和更广泛的社会群体正施加越来越大的压力，要求解决这些问题。此前构建社交媒体信息流的方法主要聚焦于推荐系统，而我们提出了一种据我们所知是新颖的、基于拍卖的信息流构建方法，其中用户对其他用户的注意力进行出价。我们的机制系统地考虑了生产者、消费者、平台运营者和社会福利。我们证明，在预算约束下，该拍卖是弱激励相容的。为平衡生产者与消费者的福利，我们在拍卖中引入了税收政策，以提高带有负外部性内容的成本。在常见社交网络拓扑结构和经实证观测的网络上进行的模拟显示，不同的信息流算法会对不同利益相关者进行差异化优先级排序。在经实证观测的网络上，我们提出的基于拍卖的机制产生的生产者福利平均比对比算法高36.3%；在合成网络上，生产者福利平均比对比算法高31.4%。在所有评估的网络类型中，我们的方法还产生了比基线方法更公平的注意力分配分布。我们提出的机制从系统层面解决了注意力分配问题，平衡了不同利益相关者的需求。
英文摘要
Social media recommendation systems create conflict: content producers, content consumers, platform operators, and social pressures struggle to direct the allocation of attention in their favor while also facing competing societal pressures. Producers may feel pressure to optimize for recommendation algorithms and consumers may be exposed to content with negative externalities such as polarization and misinformation. Platforms aim to maximize user engagement, oftentimes resulting in over consumption from consumers. There is mounting pressure from policymakers and broader society to address these issues. Prior methods for constructing social media feeds have focused on recommendation systems. Instead, we propose a, to our knowledge, novel auction based method for feed construction where users bid for the attention of other users. Our mechanism systematically considers producers, consumers, platform operators, and social welfare. We show that our auction is weakly incentive compatible under budget constraints. To balance between producer and consumer welfare, we introduce a tax policy to the auction to increase the cost of content with negative externalities. Simulations over common social network topologies and an empirically observed network show how different feed algorithms prioritize different stakeholders. Our proposed auction based mechanisms produce on average 36.3% higher producer welfare than comparison algorithms on the empirically observed network and 31.4% higher producer welfare than comparison algorithms on the synthetic networks. Our methods also produce more equitable distributions of attention than baseline methods across all evaluated network types. Our proposed mechanism addresses attention allocation at a systematic level, balancing between the needs of different stakeholders.
