Crowdsourcing over Networks: Challenges and Performance Guarantees Swaprava Nath Advisor: Prof. Y. Narahari Department of Computer Science and Automation Indian Institute of Science, Bangalore Email: swaprava@gmail.com, Web: http://swaprava.byethost7.com November 2012 Abstract Individuals and organizations often face the challenge of executing tasks for which they do not have enough resources or expertise. Outsourcing tasks to experts at a cost helps them execute it without procuring any extra resource. With the advent of the Internet, outsourcing has become even more convenient, and in particular, online social networks have given access to a huge crowd with plenty of diverse expertise. Today it is easy to find a group of people to collectively solve a problem or generate a content aggregating collective knowledge, which is known as crowdsourcing in literature. However, the design challenge is to ensure integrity of the solutions, individuals, and their strategies. My research focuses on these kinds of problems that lie in the intersection of computer science and microeconomics. In particular, I have looked into three key research challenges in crowdsourcing as depicted in Figure 1. I formulated them as optimization problems under the influence of the strategic behavior of the participants, and provided mechanism design solutions. An integral part of all these analyses is that the participants are connected over a social network, through which they can invite their friends to join a specific mission. The tasks can be outsourced to a known set of experts or to a loosely connected crowd. Each of them exhibits different strategic behaviors and thus the optimal solutions differ. My research investigates the game theoretic questions in both these domains and serves to yield certain performance guarantees while aggregating the information, skill, or effort. My theoretical study is supplemented with experimental illustration. The mathematical tools used in the analyses are game theory, optimization, stochastic processes, statistics, and decision theory. Crowdsourcing Skill Elicitation • Stochastic skill transition, known probability transition matrix, dependent values [Nath et al., 2011] • Learning the transition proba- bilities (ongoing work) Structural Manipulation • Sybilproof mechanisms, atomic tasks [Nath et al., 2012b] • Non-atomic tasks with infor- mation manipulation [Nath et al., 2012a] Efficient Team Formation • Trade-off in the information propagation and free-riding in hierarchies: [Nath et al., 2012c] • Dividing tasks to teams (ongo- ing work) Figure 1: Graphical overview of the research components of the Crowdsourcing problem. 1 Motivation and Research Objectives Aggregating the knowledge from a pool of individuals in order to perform a task has come to be popularly known as crowdsourcing . The individuals in a crowd are of varied expertise and they are often socially connected in the form of a network. The list of examples of crowdsourcing is rather long. Posting tasks on Amazon Mechanical Turk or oDesk, asking questions on Yahoo! Answers, getting user inputs for ESP game, Netflix challenge, DARPA Red Balloon and CLIQR quest challenge are all examples of task executing that leverage the collective knowledge of the crowd. My research focuses on the setting where a designer tries to efficiently execute a task with the help of a network- connected strategic crowd guaranteeing certain design goals. Examples include tagging an image or a document on Amazon Mechanical Turk, locating an object, e.g., the red balloons, or providing answer to a query in query incentive networks etc. The design goals I considered are that of eliciting the true skills of the workers, ensuring that the workers do not create fake identities in order to maximize their payoff, and to form a network hierarchy that is highly productive. 1