Recommendation algorithms play an increasingly central role in our information ecosystem. Yet, so far, they are mostly designed, parameterized and updated unilaterally by private groups or governmental authorities, based on insecure data from increasingly many fake accounts. In this paper, we present an end-to-end permissionless collaborative algorithmic governance pipeline with security guarantees, which is deployed on the open-source platform https://tournesol.app. Our pipeline has essentially four steps. First, voting rights are assigned to the contributors, based on Sybil-resilient email domains and on a novel secure trust propagation algorithm. Second, a generalized Bradley-Terry model turns contributors' pairwise alternative comparisons into scores. Third, contributors' scores are collaboratively scaled, by an adaptation of the robust sparse voting solution Mehestan. Finally, scaled scores are post-processed and securely aggregated into human-readable global scores, which are used for recommendation and display. We believe that our pipeline lays an appealing foundation for any collaborative, effective, scalable, fair, interpretable and secure algorithmic governance.

翻译：推荐算法在我们的信息生态系统中扮演着日益核心的角色。然而，迄今为止，它们大多由私人团体或政府机构基于来自日益增多的虚假账户的不安全数据，单方面设计、参数化并更新。本文提出了一种端到端的、具有安全保证的无许可协作算法治理流水线，该流水线已部署在开源平台 https://tournesol.app 上。我们的流水线主要包括四个步骤。首先，基于抗女巫攻击的电子邮件域名和一种新颖的安全信任传播算法，为贡献者分配投票权。其次，采用广义Bradley-Terry模型将贡献者对候选项目的两两比较结果转化为分数。第三，通过自适应鲁棒稀疏投票方案Mehestan对贡献者的分数进行协作缩放。最后，对缩放后的分数进行后处理，并安全聚合为人类可读的全局分数，用于推荐和展示。我们相信，我们的流水线为任何协作、高效、可扩展、公平、可解释且安全的算法治理奠定了具有吸引力的基础。