Algorithmic Advances for Fair and Efficient Decision-Making in Online Platforms

Abstract

Modern online platforms—such as recommendation systems, advertising markets and e-commerce sites—operate in dynamic and complex environments where efficient algorithmic decision-making is essential. These platforms must continuously adapt to rapidly changing user behaviors, market fluctuations, and data uncertainties while optimizing for both learning efficacy and revenue generation. However, focusing solely on performance can lead to biased outcomes and inequitable treatment of users and items, raising concerns about fairness. Balancing efficiency and fairness is therefore crucial for sustainable platform growth. In this thesis, we tackle these challenges by developing novel algorithmic frameworks and methods that integrate fairness considerations with robust learning and optimization techniques. We explore these problems from three distinct perspectives, each contributing to enhancing the decision quality and fairness considerations in online decision-making.

In Chapter 2, we first focus on the topic of efficiency, by addressing the challenge of performing online learning in a highly non-stationary environment. User behaviors and preferences often change over time, making it difficult for traditional algorithms to maintain good performance. This issue is particularly prevalent in real-world applications such as recommendation systems and advertising platforms, where shifts in user dynamics can undermine decision-making efficacy. To tackle this, we propose a novel algorithm for the widely adopted multi-armed bandit framework that enables platforms to adaptively learn in a fast-changing environment characterized by auto-regressive temporal dependencies.

In Chapter 3, we shift our focus to the realm of fairness and explore how fairness considerations can be effectively integrated into the context of assortment planning. As algorithmic recommendations become integral to platform operations, a purely revenue-driven approach can result in highly imbalanced outcomes, leading to certain items receiving minimal exposure and exiting the platform in the long run. To address this, we develop a combinatorial optimization framework that incorporates fairness constraints, ensuring equitable exposure and opportunities for all items on the platform. We design a series of polynomial-time approximation algorithms to solve the fair assortment problem. Through numerical studies on both synthetic data and real-world MovieLens data, we showcase the effectiveness of our algorithms and provide insights into the platform's price of fairness.

In Chapter 4, we bridge the topics of fairness and learning efficiency by examining how to achieve multi-stakeholder fairness in a multi-sided recommendation system. Here, the challenge is multifaceted, including ensuring high platform revenue, maintaining fair outcomes for diverse stakeholders, and enabling robust learning amidst data uncertainty. We propose a novel optimization framework that maximizes platform revenue while enforcing fairness constraints for both items and users, accommodating various fairness notions and outcome metrics. Building on this, we introduce a low-regret online learning and optimization algorithm that dynamically balances learning and fairness—two objectives that are often at odds. Finally, we demonstrate the efficacy of our approach via a real-world case study on Amazon review data and offer actionable guidelines for implementing fair policies in practice.