Friday, November 6, 2015 in Cass 104 at 3:30 pm.

John Doucette
University of Waterloo
(Candidate for Tenure track position in Computer Science)

Social Choice, Imputation and Partial Preferences: Making Better Group Decisions when Information is Hidden

Voting, or social choice, is a core component of many multiagent systems, allowing communities of agents to reach compromises or aggregate information. Voting systems based on ranked preferences can outperform simpler schemes, but have information requirements that are unrealistically high. I propose a new technique for determining the winner an election when most of the electorate is unwilling or unable to state their full preferences, using conventional machine learning algorithms. This novel “imputation-based” approach to social choice outperforms existing state of the art methods on real-world data, and provides a theoretical link between classification algorithms and social choice. Additional results include a characterization of the where the imputation-based approach will perform best, in terms of the margin of victory in a counterfactual election; a novel trajectory-based algorithm for learning preferences; and an axiomatic characterization of the fairness of different imputation methods. I will also briefly discuss other projects at the intersection of game theory, learning, and social choice.