ICES: Small: Collaborative Research: Robust Preference Aggregation

个人的喜好可能很复杂。该项目调查人们的偏好实际上是否符合CP-Net(条件偏好网络)形式，如何根据不同的投票方法将这些偏好组合在一起，以及这些方法有多容易受到操纵或“战略投票”的影响。最后，使用Netflix挑战赛中的大量偏好数据建立操纵算法效率的统计模型，并比较不同共识方法的性能。实验室实验将检查人们是否表现出可以用CP网建模的偏好。该项目将思想从贝叶斯网络中的聚合(条件概率模型)扩展到CP-网中的聚合(条件偏好模型)。团队调查了来自噪声数据的统计推断如何与战略可操纵性和贿赂相关或相互作用。该项目扩展了投票系统上的行为社会选择和计算社会选择及其操纵，从以评级、排名或子集表示的偏好到以CP网表示的偏好，研究了投票方法的性能和操纵方案在Netflix挑战数据集中的真实偏好数据上的效率。这些数据，即数十万(略感不安)的个人电影排名，是几年前出于数据挖掘的目的而发布的。人们可以根据一小部分电影的排名来提取个人的“选举”，评估和比较各种聚合方法，并对特别容易或特别有弹性的战略操纵的投票场景进行实证表征。在这项研究的更广泛的影响中，除了几个不同的研究领域跨几个科学学科的整合和交叉滋养之外，还有开发更充分的个人和集体决策工具，这将帮助个人、团体、组织和社会改善决策。

An individual's preferences can be complicated. The project investigates whether people's preferences do, in fact, conform to the CP-net (conditional preference network) formalism, how such preferences can be combined according to different voting methods, and how vulnerable those methods are to manipulation or "strategic voting." Finally, the very large corpus of preference data from the Netflix Challenge is used to build statistical models of the efficiency of manipulation algorithms and to compare the performance of different consensus methods.Laboratory experiments will check whether people exhibit preferences that can be modeled by CP-nets. The project extends ideas from aggregation in Bayesian networks (models of conditional probabilities) to aggregation in CP-nets (models of conditional preferences).The team investigates how statistical inference from noisy data relates to or interacts with strategic manipulability and bribery. The project extends behavioral social choice and computational social choice on voting systems, and the manipulation thereof, from preferences expressed as ratings, rankings, or subsets to preferences expressed as CP-nets.The team investigates the performance of voting methods and the efficiency of manipulation schemes on real preference data from the Netflix challenge data set. These data, namely hundreds of thousands of (slightly perturbed) personal rankings of movies, were released several years ago for data-mining purposes. One can extract individual "elections" based on the rankings of a small set of movies, evaluate and compare various aggregation methods and empirically characterize voting scenarios that are especially susceptible or especially resilient to strategic manipulation.Among the broader impacts of the research, beyond the integration and cross-fertilization of several distinct research areas spanning several scientific disciplines, are the development of more adequate individual and collective decision making tools, that will help individuals, groups, organizations, and society to improve decision making.