Acquiring and Aggregating Information from Strategic Sources

Abstract

This thesis considers, from a theoretical perspective, the design of mechanisms to accomplish the objective described in the title. Two cases of this problem are considered: information as represented by data points in a machine-learning context, and information as represented by beliefs in a general prediction context. While there is significant literature on the acquisition or aggregation problems as considered separately, past approaches are often inapplicable or inefficient when considering both together.

For the case of data, the thesis proposes an active procurement approach whereby data points are selectively purchased depending on their utility to the learning algorithm. A model is proposed and a purchasing scheme designed that interacts in black-box fashion with the user’s choice of learning algorithm. For a large class of problems, via a specific choice of learning algorithm, risk and regret bounds are proven as a function of budget and the “monetary difficulty” of the problem.

For the case of beliefs, the thesis proposes a theory of substitutes and complements of pieces of information. In particular, this theory is used to analyze prediction markets, which are natural and popular mechanisms for simultaneously acquiring and aggregating beliefs.

In addition, the thesis examines several additional problems involving information acquisition and aggregation in the fields of crowdsourcing and mechanism design with and without money.