Resolving Elections with Partial Preferences Using Imputation

Abstract

We propose a novel approach to deciding the outcome of elections when voters are unable or unwilling to state their complete preferences. By viewing the problem as an exercise in imputation, rather than direct aggregation, elections can be decided with an empirically supported guess of how voters would have voted, if they had complete information about the alternatives. We show that when certain classification algorithms are used to generate imputations, the process can be viewed as a form of voting rule in its own right, allowing application of existing results from the field of computational social choice. We also provide an analytical relationship linking the error rate of the classifier used with the election’s margin of victory, and extensive empirical support for the model using real-world electoral data. The described techniques both make extensive use of, and have applications throughout Multiagent Systems and Machine Learning.