随着数字时代的到来，人工智能（AI）系统已经成为人类日常生活的重要组成部分。然而，研究人员发现，现有AI系统存在算法公平性、价格歧视和信息茧房等诸多问题。为了保证AI系统在人类社会中的可靠应用，AI系统的算法治理以及在治理约束下的算法设计显得尤为重要。在实际应用中，AI系统通常由预测模型和决策模型组成，每个模型都面临独特的治理和设计挑战。例如，预测模型可能因为历史训练数据的偏差而产生分布外泛化问题和算法公平性问题。决策模型可能忽视社会影响而产生风险，如个性化定价中的价格歧视问题和推荐系统中的信息茧房效应。本文针对上述问题进行了深入探讨，主要贡献和创新包括：1、针对分布外泛化问题，本文从理论上分析了其中一类重要子问题（即协变量迁移泛化问题）的最优解。本文首先分析了在应对该任务时最小且最优的变量集合。然后本文进一步证明了基于独立性的样本加权算法能够识别出这一变量集合，从而展现该算法在处理协变量迁移泛化问题的有效性。2、针对算法公平性问题，本文引入了条件公平概念以同时实现因果解释性和可计算性的目标。传统群体公平概念缺乏因果解释性，而基于因果的公平则依赖于强假设，不具有可计算性和可扩展性。因此，本文提出了全新的条件公平概念以结合上述两种公平性概念的优势。同时，本文还设计了一种新的正则化器，该正则化器可以有效平衡算法的准确性和条件公平性。3、针对价格歧视问题，本文提出了个性化定价的两个监管约束并证明了它们在监管价格歧视上的有效性。本文研究了垄断企业价格歧视场景，从理论上分析了所提出的监管约束对社会福利的潜在影响。理论结果表明，在常见的市场需求分布中，所提出的政策约束能够有效地平衡消费者和生产者的利益。4、针对信息茧房问题，本文从消费者和内容创作者两个角度探讨了推荐系统中信息茧房的形成。从消费者角度，本文提出了一个全新的推荐模型，其目标是在破除信息茧房的同时提升平台的长期收益。从内容创作者角度，本文利用算法博弈论分析了创作者之间的竞争行为，证明这种竞争行为会降低社会总收益并导致信息茧房的形成。
The advent of the digital era has elevated artificial intelligence (AI) systems to a pivotal position, significantly influencing our daily lives. However, researchers have identified several critical issues when applying AI systems, including algorithmic fairness, price discrimination, and the information cocoon. To ensure AI‘s reliable and trustworthy application, it is important to study the governance of AI systems and the design of AI under regulatory constraints. In practice, AI systems are typically comprised of predictive models and decision-making models, each facing distinct governance challenges. For example, predictive models often contend with biased data, leading to issues in out-of-distribution generalization and algorithmic fairness. Conversely, the governance of decision-making models primarily focuses on assessing the societal impacts of AI systems, including price discrimination in personalized pricing and the information cocoon effect in recommender systems. This dissertation studies the above issues and the main contributions and innovations of this dissertation include the following:1. For the out-of-distribution generalization problem, this dissertation studies an important sub-problem (i.e., the covariate-shift generalization problem) and analyzes the optimal solution to this sub-problem. This dissertation first identifies a variable set, which is the optimal and minimal variable set to address the covariate-shift generalization problem. Based on the variable set, this dissertation further proves the effectiveness of independence-driven importance weighting algorithms by showing that this algorithm could identify the variable set.2. For the algorithmic fairness problem, this dissertation proposes a novel conditional fairness notion to ensure causally explainability and computability. Traditional group fairness metrics, though prevalent, lack explainability. Although causality-based fairness offers a solution to this limitation, it requires strong assumptions about causal structures and lacks scalability and computability. Therefore, this dissertation proposes a new conditional fairness notion to combine the advantages of group fairness and causality-based fairness notions. In addition, this dissertation also proposes a novel regularizer that could effectively achieve a balance between models‘ performance and conditional fairness.3. For the price discrimination problem, this dissertation proposes two regulatory policies and proves their effectiveness in regulating personalized pricing. This dissertation studies the monopoly setting and theoretically analyzes the effects of the proposed policies on social welfare. Theoretical analysis suggests that under common demand distributions, the proposed regulations can efficiently balance the benefits of consumers and producers.4. For the information cocoon problem, this dissertation analyzes the formation of the information cocoon issue in recommender systems through the lens of both consumers and content creators. From the consumers‘ perspective, this dissertation further proposes new models for efficiently mitigating the information cocoon problem and potentially enhancing the long-term utility of recommender systems. From the content creators‘ perspective, the dissertation employs algorithmic game theory to analyze competitive behaviors of content creators, demonstrating that such competition decreases social welfare and contributes to the formation of information cocoons.
