A House Divided Against Itself? Dividing Crowds to Improve Analysis & Design of Crowd-Based Activities
In the past several years, crowd-activities have expanded dramatically in both their scope and their size – more diverse activities are involving online crowds, and the number of people (and money) involved in such activities is growing fast. I will focus in this talk on 2 topics, which share as a central feature the partitioning of the crowd into sub-groups.
In peer-selection, agents try to select the top k agents among themselves. We present several algorithms that try to make this process impartial, i.e., agents will not benefit from lying about their peers, while still choosing "good" agents. We will show algorithms that provably reach impartiality using the division of the agents into subgroups, and then show a set of simulations that give us further insight into which techniques work better in scenarios resembling real-world settings.
From there, we will move to district selection: decision making in settings where each partition (a company sub-unit, an electoral district, etc.) reaches a decision, and the decisions from all partitions are amalgamated to a final choice. We explore how this effects representability, and in the case of geographical manipulations ("gerrymandering"), present both computational hardness results and experiments on real-world data.