There are many ways one could identify photos that a person might like: analyzing the content of photos to look for similar features, identifying tags in a person's photos and looking for similarly-tagged photos, using friends' photos or preferred photos, etc. In a research project with Flickr, we created an experiment that compares several possible means of recommending photos. If you are an active Flickr user, you can go to koala.sandbox.yahoo.net and try it out.
Amazon's Mechanical Turk is a crowdsourcing platform where any individual can come and do micro-tasks such as labeling images or classifying websites for small amounts of money. As an online labor market, it also can serve as a pool of low-cost paid participants for behavioral research. Based on our accumulated experiences conducting studies on Mechanical Turk, Sid Suri and I wrote a guide on how to use the platform for behavioral research.
Research has shown that people tend to express their unconscious, automatic attitudes through unconscious, nonverbal behavior. For instance, the Implicit Attitude Test demonstrates that people's reaction times are guided by these implicit attitudes. Other research has shown that people are influenced by nonverbal behaviors. This project ties these two lines of research together, and show that one person's implicit attitudes can influence another person's attitudes.
Most research using the Implicit Attitude Test relies on offline computer programs or proprietary software. At one time this was necessary, because the timing provided by browsers was insufficient to handle millisecond timing with a reasonable amount of accuracy. These days, however, computers and browsers are fast enough to handle very brief timing with javascript. To this end I designed an open source IAT (available here) that uses a simple text file to define the inputs to the IAT and runs easily on a server running PHP.
There are many theories about how culture spreads between individuals. Studying the phenomenon in the past has relied on population-level metrics (e.g., Rogers, 1995) or laboratory experiments of dyadic influence (e.g., Sherif, 1935). However, recent technologies provide a new and unique glimpse into the real-world process. On the social media site Twitter, media sites, bloggers, corporations, and regular people create a network for sharing information, often in the form of URLs. We observe the diffusion of URLs on this micro-blogging site and consider the factors that lead to the spread of a URL.
There are a number of reasons why one's position in their social network could improve their success. By having access to more people, one has access to more information and more opportunities. By bridging different groups, one has access to different types of people and information. By belonging to a tightly-knit group, one has social support and community. The subject of this research is determining which of these structural features, or combination of these features, affect an individual's success. We look at citations received in co-authorship networks, comments received in blog networks, and trade between industries.
In many studies of real-world social networks, especially communication networks, researchers make an arbitrary decision about what constitutes a edge. However, this decision has important consequences for the conclusions that are reached about the nature of the network. In this work we show that the features of the network can change radically depending on the threshold you choose as sufficient strength for the tie. In addition, we present a method for finding the optimal threshold given a particular prediction problem.
Many new instances of "peer production" have appeared on the web, from Wikipedia, to Digg, to Amazon's Mechanical Turk. Some researchers have begun to explore who is contributing and why, but understanding the relationship between incentives, motivations, and the efficacy of the peer production systems has largely gone unexplored. By manipulating the pay rate and task features, and measuring both the quantity and quality of tasks completed, we found that participants responded to pay by increasing the number of tasks completed, but did not change the quality of work performed. One potential explanation comes from the results of a post-task survey, in which participants consistently reported they felt the value of the task was slightly more than they earned.
Although homophily is a well-studied phenomenon, especially with respect to observable characteristics such as race and eduation, empirical studies disagree over how much diversity of opinions exists within local social networks, and how much awareness individuals have of their neighbors' views. We created an application on Facebook in which participants were asked about their own attitudes and their beliefs about their friends' attitudes. Although we found considerable attitude homophily, the results show that friends disagree frequently, and considerably more than they think they do. In particular, friends are typically unaware of their disagreements even when they say they talk about politics. It appears that most respondents' understanding of their friends' attitudes is not based on knowledge obtained through discussion, but instead mostly from relying on their knowledge of the attitudes' prevalence in the population and anchoring off of their own views.
In this study we experimentally manipulated the communication networks between individuals searching a problem space. Individuals picked a number between 0-100 which returned a score based on a continuous fitness function. They got to see their score and the score of their neighbors in the communication network. In this way, numbers that received good scores good diffuse through the group. We found that different network structures performed better or worse (got higher or lower average scores) depending on the shape of the fitness function. Specifically, communication structures that had larger diameters performed better in problem spaces that required more exploration, such as multimodal functions, while networks with short paths did better in simple problem spaces. This suggests that some social network structures might be better for different kinds of group problem solving tasks.
Berger, Heath, and Ho (2007) posited that tastes are used by group members to signal group identity, and that sometimes these tastes are poached by posers who want to appear like another group. We explored an agent-based model that simulated this process, which demonstrated this poaching is sufficient to cause cultural cycles, but the effect can be mediated by costly tastes.