Details of talks (speaker, title, abstract, videos) will be updated here.
Subsidizing Sequential Search
Speaker:
Nicole Immorlica
Abstract:
We study markets where firms compete for consumer attention by subsidizing costly product inspection. Such subsidies do not alter product quality, but they reshape the consumer's search order by lowering inspection costs....
We study markets where firms compete for consumer attention by subsidizing costly product inspection. Such subsidies do not alter product quality, but they reshape the consumer's search order by lowering inspection costs. We establish a subsidy–sorting principle: in any equilibrium, higher–quality firms offer weakly larger subsidies, inducing consumers to search in descending subsidy order. Under the Intuitive Criterion, a unique equilibrium survives: low–quality firms do not subsidize, intermediate types separate with strictly increasing subsidies, and high–quality firms pool by offering full subsidy. This outcome minimizes consumer's search costs among all equilibria and implements efficient inspection. We then extend the analysis to AI mediated platforms that can mint and price “inspection tokens”. The platform’s optimal linear pricing induces excessive inspection relative to the social optimum. While this distortion does not reduce consumer welfare, it reallocates sellers' surplus toward the platform.
More
Video: TBA
A Characterization of Strategy-Proof Probabilistic Assignment Rules
Speaker:
Souvik Roy
Abstract:
We study the classical probabilistic assignment problem, where finitely many indivisible objects are to be probabilistically or proportionally assigned among an equal number of agents...
We study the classical probabilistic assignment problem, where finitely many indivisible objects are to be probabilistically or proportionally assigned among an equal number of agents. Each agent has an initial deterministic endowment and a strict preference over the objects. While the deterministic version of this problem is well understood, most notably through the characterization of the Top Trading Cycles (TTC) rule by Ma (1994), much less is known in the probabilistic setting. Motivated by practical considerations, we introduce a weakened incentive requirement, namely SD-top-strategy-proofness, which precludes only those manipulations that increase the probability of an agent’s top-ranked object. Our first main result shows that, on any free pair at the top (FPT) domain (Sen, 2011), the TTC rule is the unique probabilistic assignment rule satisfying SD–Pareto efficiency, SD–individual rationality, and SD–top-strategy-proofness. We further show that this characterization remains valid when Pareto efficiency is replaced by the weaker notion of SD–pair efficiency, provided the domain satisfies the slightly stronger free triple at the top (FTT) condition (Sen, 2011). Finally, we extend these results to the ex post notions of efficiency and individual rationality.
More
Video: TBA
Reward Schemes in Blockchain Environments: Shapley Value and Proportional Sharing
Speaker:
Vangelis Markakis
Abstract:
This talk focuses on a model for pool formation in Proof of Stake protocols. In such systems, stakeholders can form pools as a means of obtaining regular rewards from participation in ledger maintenance and block production, with the power of each pool being dependent on its collective stake....
This talk focuses on a model for pool formation in Proof of Stake protocols. In such systems, stakeholders can form pools as a means of obtaining regular rewards from participation in ledger maintenance and block production, with the power of each pool being dependent on its collective stake. The question of interest here is the design of reward schemes that suitably split earned profits among pool members. With this in mind, we initiate a study of the well known Shapley value scheme into the context of blockchains. We provide comparisons between the Shapley mechanism and the more standard proportional scheme, in terms of attained decentralization (i.e., number of pools formed at equilibrium), and in terms of susceptibility to Sybil attacks, i.e., the strategic splitting of a players' stake with the intention of participating in multiple pools for increased profit.
More
Video: TBA
Individual Agency Over Algorithmic Risk Prediction
Speaker:
Vijay Keswani
Abstract:
Algorithmic risk assessment tools perform the task of predicting the context-relevant risk of individuals obtaining certain outcomes. While increasingly prevalent in real-world applications, their usage becomes controversial...
Algorithmic risk assessment tools perform the task of predicting the context-relevant risk of individuals obtaining certain outcomes. While increasingly prevalent in real-world applications, their usage becomes controversial when the predicted risk scores fail to be actionable, i.e., when one can’t design suitable interventions to positively impact the risk scores. We explore this issue by modeling the “agency” associated with attributes used for algorithmic risk predictions. By comparing the relative agency individuals have over attributes used for risk prediction to their predictive relevance, we formalize an audit procedure to assess the actionability of risk prediction in practice.
More
Video: TBA
Towards a Theory of Equilibrium in Data Markets
Speaker:
Bhaskar Ray Chaudhury
Abstract:
The algorithmic study of market equilibria has been a cornerstone of economics and computation since its inception, with rich theories developed around equilibrium and stability notions such as Nash, Stackelberg, and Competitive Equilibria...
The algorithmic study of market equilibria has been a cornerstone of economics and computation since its inception, with rich theories developed around equilibrium and stability notions such as Nash, Stackelberg, and Competitive Equilibria. In this talk, I explore how these classical concepts extend—or fail to extend—to a new and increasingly central economy: the data economy. With the increasing integration of AI-ML technologies in the industry, data has emerged as one of the most valuable assets of the 21st century. However, unlike traditional goods, data is non-rival—it can be freely duplicated and shared without depletion. This fundamental property challenges standard equilibrium frameworks that assume scarcity (limited supply) and exclusivity (no two agents can simultaneously benefit from a resource). We introduce a formal model of data markets, explicitly capturing (1) the role of data in improving predictive performance as data buyers’ utility functions, and (2) the implications of data’s non-rival nature for equilibrium concepts. Building on this model, we examine the existence, geometry, and computational aspects of various equilibrium notions, highlighting key parallels and departures from traditional markets. The talk will begin with a brief tutorial on equilibrium concepts—no prior background required—and conclude with open problems and emerging directions for the theory and design of data markets.
More
Video: TBA
Learning in the Presence of Strategic Agents
Speaker:
Ganesh Ghalme
Abstract:
Humans have come to rely on machines to make decisions concerning their welfare, with applications in criminal risk assessment, health premium calculations, screening resumes, and processing loan applications...
Humans have come to rely on machines to make decisions concerning their welfare, with applications in criminal risk assessment, health premium calculations, screening resumes, and processing loan applications when these learning-based decision-making systems interact with humans, they present interesting challenges. For instance, when the decision rule is known, rational agents may respond to it by manipulating their features to obtain favourable outcomes; for example, by paying off outstanding loans to be eligible for a new loan. In such a case, the goal is to find a strategy-robust decision rule. In this talk, I will present a strategic classification framework in which agents misrepresent their features at test time to game the system's classifier to achieve a favorable outcome. We will also explore a partial information (in-the-dark) setting where agents have access to a rule via previous decisions made by the classification rule (such as loan decisions made for acquaintances) and attempt to game the reconstructed rule from available information.
More
Video: TBA
Fairness and Incentives in Data Sharing Platforms
Speaker:
Ruta Mehta
Abstract:
With the explosion of data and data-driven technologies, the design of principled data-sharing paradigms has become increasingly important for both economic and societal welfare. Federated learning (FL) offers a powerful framework for leveraging rich, distributed datasets while...
With the explosion of data and data-driven technologies, the design of principled data-sharing paradigms has become increasingly important for both economic and societal welfare. Federated learning (FL) offers a powerful framework for leveraging rich, distributed datasets while preserving data privacy. Yet the inherent heterogeneity of these datasets raises fundamental challenges in defining and ensuring fairness among participating agents, leading to potential incentive misalignments. For example, without appropriate compensation, an agent contributing high-quality data may have little motivation to participate when others’ data are of lower quality. Moreover, due to privacy costs, communication overheads, and competitive risks, agents may not be incentivized to share their high-quality data at all. In this talk, I take a social choice and game-theoretic perspective on these fairness and incentive challenges. I will show how tools from social choice theory (SCT) and federated learning (FL) can inform and strengthen one another, yielding new conceptual insights and opening up promising research directions.
More
Video: TBA