AF: Small: Learning and Optimization with Strategic Data Sources

Harvard University, Cambridge MA

Investigators

Abstract

The goal of this research project is to develop new results in machine learning and optimization when training data for machine learning or information about optimization problems is acquired from strategic sources. We are blessed with unprecedented abilities to connect with people all over the world: buying and selling products, sharing information and experiences, asking and answering questions, collaborating on projects, borrowing and lending money, and exchanging excess resources. These activities result in rich data that scientists can use to understand human social behavior, generate accurate predictions, find cures for diseases, and make policy recommendations. Machine learning and optimization traditionally take such data as given, for example treating them as independent samples drawn from some unknown probability distribution. However, such data are possessed or generated by people in the context of specific rules of interaction. Hence, what data become available and the quality of available data are results of strategic decisions. For example, people with sensitive medical conditions may be less willing to reveal their medical data in a survey and freelance workers may not put in a good-faith effort in completing a task. This strategic aspect of data challenges fundamental assumptions in machine learning and optimization. The research project takes a holistic view that jointly considers data acquisition with learning and optimization. It will bring improved benefits in business, government, and societal decision-making processes where machine learning and optimization are widely applicable. The research project also involves the mentoring of PhD students, innovation in graduate teaching, and engagement of members of underrepresented groups in research. The PI will pursue a broad research agenda developing a fundamental understanding of how acquiring data from strategic sources affects the objectives of machine learning and optimization. The first set of goals aims to develop a theory for machine learning when a learning algorithm needs to purchase data from data holders who cannot fabricate their data but each have a private cost associated with revealing their data. A notion of economic efficiency for machine learning will be established. The second set of goals will further advance the frontier of machine learning by designing joint elicitation and learning mechanisms when data are acquired from strategic agents but the quality of the contributed data cannot be directly verified. The third set of goals will develop optimization algorithms with good theoretical guarantees when parameters of an optimization problem may be unknown initially but the algorithm designer can gather information about the parameters from strategic agents.