EAGER: Fairness-Aware Personalized Recommendations

这个项目的目标是创建有效的信息管理员推荐模型，该模型可以为个人用户个性化，同时保持重要的公平属性。信息馆长作为高质量精选内容的渠道，提供独特的专业知识、决策的可信度和获取新内容的途径。这种多样性和由此产生的异质性--在内容类型、社会关系、策展人的动机等方面--对有效的个性化提出了极大的要求。此外，现有的以点赞、关注关系和其他互动的形式表达的对这些策展人的偏好通常是稀少的。因此，个性化馆长推荐的一个关键挑战是解决稀疏性问题，同时在复杂、噪声和异质环境中仔细建模馆长。使这一挑战雪上加霜的是，目前大多数对信息馆长的访问是通过集中式平台(如搜索引擎、社交网络和传统新闻媒体)进行的，这意味着个人偏好可能与这些平台的目标不一致，从而导致对馆长的潜在偏见(甚至是有限的)访问。一个关键问题是如何在馆长推荐中维护公平性。该项目的预期成果包括可以对现有网络和社交媒体平台产生积极影响的研究进展，以及为未来信息精选推荐的进展提供理论基础。在大规模发现信息管理人、可靠地将用户与适当的管理人联系起来以及确保这类管理人的公平属性方面取得的进展，对于支持知情民众的可信信息至关重要。通过将这些研究进展、数据集和工具包带到更广泛的研究社区，该项目可以从其他研究人员的互补努力中刺激更多的进展。此外，这个项目将开发新的课堂材料，新的外展努力，以及新的扩大参与的研讨会和研讨会。本项目将探索和测试四个具有挑战性的研究问题：(1)异质环境中基于学习张量的推荐器。由于用户对馆长的偏好可能会受到许多上下文因素的影响，因此第一个任务将直接将用户、馆长、主题和其他因素之间的多种不同关系直接纳入基于张量的方法中。(2)馆长推荐的神经个性化排序。作为这种基于张量的方法的补充，该项目还将探索个性化馆长推荐的新神经模型的能力。与基于张量的方法相比，神经模型有望在模型设计中提供更大的灵活性，通过激活增加非线性，并提高性能。(3)神经+张量混合模型。第三，这个项目将研究新的混合模型，它结合了基于张量的方法(原则性的和可解释的)和基于神经的方法(承诺改善性能)的优点。(4)公平意识学习。最后，本项目将探索公平感知约束下的个性化推荐。由于推荐者可能会从用于优化他们的培训数据以及平台目标和个人偏好之间的不一致中继承偏见，该项目将构建新的公平意识算法，通过增强主题、策展人和观点的多样性来增强用户的能力。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The goal of this project is to create effective information curator recommendation models that can be personalized for individual users, while maintaining important fairness properties. Information curators serve as conduits to high-quality curated content, providing unique specialized expertise, trustworthiness in decision-making, and access to novel content. This variety and the resultant heterogeneity -- in terms of content types, social relations, motivations of curators, etc. -- place great demands on effective personalization. Further, existing expressed preferences for these curators in the forms of likes, following relationships, and other interactions are often sparse. Hence, a key challenge for personalized curator recommendation is tackling sparsity while carefully modeling curators in complex, noisy, and heterogeneous environments. Compounding this challenge, most current access to information curators is mediated by centralized platforms (like search engines, social networks, and traditional news media), meaning that personal preferences may not align with the goals of these platforms, leading to potentially biased (or even limited) access to curators. A key question is how to maintain fairness properties in curator recommendation. The expected results of this project include research advances that can positively impact existing web and social media platforms, as well as provide a theoretical foundation for future advances in information curation recommendation. The advances in uncovering information curators at scale, reliably connecting users to appropriate curators, and ensuring fairness-preserving properties of such curators are critical for trustworthy information supporting an informed populace. By bringing these research advances, datasets, and toolkits to the wider research community, this project can spur additional advances from complementary efforts by other researchers. Further, this project will develop new classroom materials, new outreach efforts, and new broadening participation workshops and seminars.This project will explore and test four challenging research problems: (1) Learning Tensor-Based Recommenders in Heterogeneous Environments. Since user preferences for curators may be impacted by many contextual factors, this first task will directly incorporate the multiple and varied relationships among users, curators, topics, and other factors directly into a tensor-based approach. (2) Neural Personalized Ranking for Curator Recommendation. Complementary to such a tensor-based approach, this project will also explore the capabilities of new neural models of personalized curator recommendation. Neural models promise potentially more flexibility in model design, added nonlinearity through activations, and improved performance relative to tensor-based approaches. (3) Hybrid Neural+Tensor Models. Third, this project will investigate new hybrid models that combine the benefits of tensor-based methods (which are principled and interpretable) with neural-based methods (which promise improved performance). (4) Fairness-Aware Learning for Curation. Finally, this project will explore personalized recommendation under fairness-aware constraints. Since recommenders may inherit bias from the training data used to optimize them and from mis-alignment between platform goals and personal preferences, this project will build new fairness-aware algorithms that can empower users by enhancing diversity of topics, curators, and viewpoints.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.