The usage of data is ubiquitous and still growing. As machine learning technologies are adopted, people who have a stake in the output of these algorithms will change their behavior. We already see this today. For example, energy consumers will inflate their baseline consumption to trick utility companies when enrolling in direct load control programs. Disgruntled residents will report non-existent accidents to try and change how traffic is routed. In this talk, I will present our work characterizing the game-theoretic equilibria of data markets with strategic data sources, as well as some experimental results where participants shared data with a social choice mechanism.
For the first part of the talk, we model the data market between data sources and data aggregators. In our model, data sources are effort-averse and data aggregators purchase data from data sources. We show that under this model, data aggregators will free-ride and cause social inefficiencies. Additionally, there is non-uniqueness of the generalized Nash equilibria, which has implications for the stability of data markets. For the second part of the talk, we ran a 5-month long experiment where we controlled the lights in the office space of 27 employees. The light's brightness was determined by a modified Vickrey-Clarke-Groves mechanism. In theory, these mechanisms are famous for implementing the socially desirable outcome and incentivizing truthful reporting. In contrast, our experimental results show that participants were significantly less satisfied with light brightness and incentives in our mechanism over time, which highlights design requirements for the deployment of Internet of Things technologies in shared settings, where the opinions of multiple people must be reconciled.
