TUES. & WED. / IDEAL/CS Seminar
September 6-7, 2022 / 11:00 am CDT (Chicago Time) in Mudd Library Room 3514

For Tuesday:

Title: Machine Learning in Digital Marketplaces: Interactions between Learners, Consumers, and Producers
Speaker: Meena Jagadeesan, University of California, Berkeley
Panopto: https://www.ideal.northwestern.edu/special-quarters/fall-2022/workshops/panopto-9-6-22

Abstract: Machine learning is typically analyzed as an isolated system. In reality, machine learning is often performed by a platform that is situated within a marketplace of consumers, producers, and other platforms performing machine learning. In this talk, we examine how the learning algorithm impacts and is affected by the marketplace, and we use recommender systems as a running example.

First, we show that competition between digital platforms need not align market outcomes with user preferences, challenging conventional economic wisdom. Roughly speaking, the number of users on the platform impacts how much data the platform has and thus the quality of its service. The platform thus doesn’t have to fully cater to users to have an edge over competitors. Moreover, this misalignment with user preferences persists even if platforms are required to share their data.

Next, we examine how algorithmic recommendations shape the content that producers create. When the platform’s recommendations are personalized, producers compete to have their content shown to individual users. This fine-grained competition can incentivize producers to specialize their content, which leads to the formation of genres. Furthermore, when there is sufficient specialization, producers can achieve positive profit at equilibrium, which is indicative of monopoly-like behavior.

More broadly, we highlight several avenues through which the learning algorithm and the marketplace interact with each other. To design effective learning algorithms and regulatory policy, researchers need to formalize and account for how machine learning is affected by the marketplace in which it occurs.

Based on joint work with Nikhil Garg, Moritz Hardt, Nika Haghtalab, Michael I. Jordan, Celestine Mendler-Dünner, and Jacob Steinhardt.

 

For Wednesday:

Title: Learning Equilibria in Matching Markets from Bandit Feedback
Speaker: Meena Jagadeesan, University of California, Berkeley
Panopto: https://www.ideal.northwestern.edu/special-quarters/fall-2022/workshops/panopto-9-7-22

Abstract: Two-sided matching platforms must find market outcomes that align with user preferences while simultaneously learning these preferences from data. In economics, alignment with user preferences is captured by the notion of stability (Gale and Shapley, 1962; Shapley and Shubik, 1971); however, stability is of limited value in the learning setting, given that preferences are inherently uncertain while they are being learned.

To bridge this gap, we develop a framework for learning stable market outcomes under uncertainty, casting learning as a stochastic multi-armed bandit problem. Our primary setting is matching with transferable utilities, where the platform both matches agents and sets monetary transfers between them. Since stability is impossible to achieve under uncertainty, we instead propose an incentive-aware regret notion that captures how far a market outcome is from stable. Using this regret notion, we analyze how the structure of user preferences affects the complexity of learning. Algorithmically, we show that “optimism in the face of uncertainty,” the principle underlying many bandit algorithms, applies to a primal-dual formulation of matching with transfers and leads to near-optimal regret bounds.

Our work takes a step toward learning market outcomes that are aligned with user preferences in large, data-driven marketplaces.

Appeared at NeurIPS 2021; based on joint work with Alexander Wei, Yixin Wang, Michael I. Jordan, and Jacob Steinhardt.

Biography: Meena Jagadeesan is a 3rd year PhD student in computer science at UC Berkeley advised by Michael I. Jordan and Jacob Steinhardt. She works on the theoretical foundations of machine learning and algorithmic decision-making. Previously, she received an A.B. and S.M. in computer science from Harvard. Her research is supported by an Open Philanthropy AI Fellowship and a Paul and Daisy Soros Fellowship.