Lunch will be served at 11:45 AM.
Algorithms are increasingly deployed in personal and societal settings, relying on data generated by individuals and organizations with various (often conflicting) incentives. However, the very data that power these algorithms can also lead to unexpected harm for users and society. In this talk, I will discuss my work in this emerging field of data markets and socially aware algorithms, highlighting how market design, privacy considerations, product recommendations, and incentive structures shape the impact of algorithmic decision-making.
In the first part, I will develop a framework to study data acquisition from privacy-sensitive, strategic agents for a learning task. I will study this framework in the context of estimating a population mean (by pooling users' data) and formulate it as a Bayesian-optimal mechanism design problem. Here, an individual may share their data in exchange for compensation yet faces a private—and heterogeneous—privacy cost measured via differential privacy. I will prove that a properly designed linear estimator achieves minimax optimality. I will then describe an efficient algorithm that jointly determines the payment scheme and the estimator, balancing accuracy and privacy in an incentive-compatible way.
Besides intrinsic privacy concerns and price discrimination, I argue in the second part that the abundance of data—and the ability to customize user experiences—raises another issue: behavioral manipulation. Concretely, I will develop a continuous-time strategic experimentation model in which a platform dynamically offers multiple products (at certain prices) to a user who can gradually learn each product's quality. However, product signals can be distorted by superficial attributes --"glossiness." AI tools allow platforms to estimate glossiness more precisely, thus shaping user decisions and welfare. Using a Gittins index characterization of the platform's optimal decision, I will show that when glossiness is short-lived, the platform's superior information benefits users. In contrast, when glossiness is long-lived, the platform's superior information harms users as it enables behavioral manipulation—steering choices toward glossy, low-quality products.
Azarakhsh Malekian is a Professor of Operations Management and Statistics at the University of Toronto and is currently a Visiting Faculty Researcher at Google. Her research interests are data markets, mechanism design, algorithm design, algorithmic game theory, and networks. Her work has been recognized with several honors, including the Best Student Paper Award at the Conference on Web and Internet Economics (WINE), the University of Toronto Connaught New Research Award, the Roger Martin Excellence in Research Award at Rotman, and the Rotman Teaching Award. She currently serves as an Associate Editor for Operations Research and Management Science Journals and chairs the INFORMS Lancaster Prize Committee. She previously co-chaired the Conference on Web and Internet Economics (WINE) in 2022 and co-chaired the Economics, Monetization, and Online Markets track at The Web Conference (WebConf) in 2023. She holds a B.Sc. in Computer Engineering from Sharif University of Technology and a Ph.D. in Computer Science from the University of Maryland. Prior to joining the University of Toronto, she was a postdoctoral scholar at MIT.