Understanding how people make decisions in complex settings is crucial in many application areas, including marketing, intelligence analysis, and political decision making. Traditionally, human decision-makers are modeled as rational agents seeking to maximize some mathematical measure of utility. In fact, however, people are overwhelmed in information-rich environments, and have developed cognitive and emotional strategies to navigate such environments. One such strategy is "motivated reasoning", where information is first evaluated subconsciously for emotional content, with the goal of maintaining an existing emotional commitment, and cognitive processing of the information is then conditioned on this emotional evaluation. Political scientists have demonstrated motivated reasoning in evaluation of both candidates and issues. In general, decision-making and information-gathering are strongly influenced by emotion, prior knowledge, and the social communities to which a person belongs. Evidence suggests that accurate models of human decision-making must be complex enough to model not only utility, but prior knowledge and beliefs, human cognitive abilities, and social context. Building such cognitive models requires substantially extending the state-of-the-art in machine learning.
In the past, the ability of researchers in political psychology to develop such complex models of decision-making was limited by the amount of data obtainable obtain from surveys or human-subject experiments. The recent explosion of on-line political communities provides an opportunity to overcome this limitation. We will model human behavior for socially-driven information gathering and decision-making tasks - specifically for political decisions - by combining human-subject experiments with analysis of large datasets of social media and social interactions.
In the past, people were often forced to make decisions with too little information. Today, the converse is true: often people need to make decsions with too much information---more information than they can reasonably process in a timely way. One example of this task is political decision-making, for example, chosing what candidate to vote for in an election, given the current deluge of relevant news stories and other information. In this grant, a group of political scientists and computer scientists studied this problem in detail. As is true in many domains, decision-making in politics is influenced by emotion, and this sometimes leads to counter-intuitive behavior: for instance, sometimes negative information about a preferred candidate will actually make a voter feel more positive toward that candidate. We explored a number of hypotheses about why such effects might be observed. One conjecture is that our emotional responses in what should (in principle at least) be a rational decision-making process are not a bug but a feature: for instance, counter-intuitive behavior might be a reaction to a perceived attempt by the information provider to manipulate voters. To explore these questions,we developed a social-experiment software environment called Dynamic Process Tracing Environment (DPTE) to run experiments to test theories of information processing and emotions, including the role of social cues - the likes, dislikes, and comments the internet now allows us to share with our social networks and beyond. Using this software tool we conducted "mock elections" under different conditions, where people were given large amounts of information about imaginary candidates and asked to make choices. The setup allowed us to ask (and answer) questions such as: do people react differently to the same information if it is provided by different people? what sort of information do people tend to share with others, and how do different strategies that people use affect how others react to their information? We also conducted similar experiments (to the extent it was possible) on political tweets. In the experiments we found that counter-intuitive behavior (of the sort described above) is indeed correlated with suspicion of the source of information. We also found that people are more likely to share information that evokes emotion, especially negative emotion, than to share emotionally neutral information, and that people share more items from in-party sources than out-of-party sources. We also found that when social cues are present on information, they greatly influence how that information is processed, including creating more congnitive engagement, directing information search, and leading people to sometimes do searches that essentially confirm pre-existing preferences.
