PhD Proposal: Learning with Incomplete Information

Kiarash Banihashem
05.20.2024 13:00 to 15:00

IRB 5105

We explore the impact of incomplete data on classical machine learning problems. For many classical problems in machine learning, we often do not have direct access to the input data necessary to correctly solve the problem. For multi-agent settings, this can often happen because the information needs to be obtained from self-interested agents which behave strategically to maximize their own reward. The core issue here is misalignment of incentives as each data provider is focused on their own interest, rather than considering the social good. We study this issue in two related settings based on recommendation systems and online auctions. Even in the absence of strategic agents, in some problems the data is fundamentally unformed in that it is not fully revealed until a later time when the model is deployed. To study this challenge, we consider the classical nearest neighbors model in a contextual setting where the underlying data can change based on some context vector (e.g., a representation of the user).