Experimental Studies of Human Behavior in Social Computing Systems

Abstract

Social computing systems, fueled by the ability of the Internet to engage millions of individuals, have redefined computation to include not only the application of algorithms but also the participation of people. Yet, the true impact of social computing in the future depends on a systematic understanding of how to design interventions that produce desirable system-wide behavior. Behavioral experiments, with their fundamental ability to study causality, are an important methodology in reaching this goal.

This dissertation presents several examples of how novel experimental approaches to studying social computing systems can not only improve the design of such systems, but improve our understanding of human behavior. We investigate the differences in motivation and effects of incentives in a crowdsourcing task across paid and volunteer crowdsourcing systems, finding that varying financial incentives even at the same wage can be used to implicitly invoke different biases. We discover the effects of assembling teams of different size in a social, collaborative crisis mapping task, finding that larger groups exert less effort per individual, but make up for this loss in their ability to coordinate. We investigate the accuracy of voting in a human computation setting and how statistical models can be used to discover patterns of decision making across a population of individuals. Finally, we present the design and implementation of TurkServer, an software system that enabled these experimental studies.

The work in this dissertation suggests that experiments in social computing present an opportunity both for understanding design factors of social computing systems and for developing generalizable models of human behavior—and ultimately better theories of how people communicate and interact in our interconnected world.