Real-World Behavior Analysisthrough a Social Media LensMohammad-Ali Abbasi†, Sun-Ki Chai‡, Huan Liu†, Kiran Sagoo‡†Computer Science and Engineering, Arizona State University‡Department of Sociology, University of Hawai‘[email protected],[email protected],[email protected],[email protected]. The advent of participatory web has enabled informationconsumers to become information producers via social media. This phe-nomenon has attracted researchers of different disciplines including socialscientists, political parties, and market researchers to study social mediaas a source of data to explain human behavior in the physical world.Could the traditional approaches of studying social behaviors such assurveys be complemented by computational studies that use massiveuser-generated data in social media? In this paper, using a large amountof data collected from Twitter, the blogosphere, social networks, andnews sources, we perform preliminary research to investigate if humanbehavior in the real world can be understood by analyzing social mediadata. The goals of this research is twofold: (1) determining the relativeeffectiveness of a social media lens in analyzing and predicting real-worldcollective behavior, and (2) exploring the domains and situations underwhich social media can be a predictor for real-world’s behavior. We de-velop a four-step model: community selection, data collection, online be-havior analysis, and behavior prediction. The results of this study showthat in most cases social media is a good tool for estimating attitudesand further research is needed for predicting social behavior.1 IntroductionThe advent of participatory web has created user-generated data [1], that leavemassive amounts of online “clues” that can be examined to infer the attributesof the individuals who produced data. As it becomes easier and easier to createcontent in the virtual world, more and more data is generated in various aspectsof life for studying user attitudes and behaviors. Sam Gosling in [7] reveals howhis team gathers a large amount of information about people without asking anyquestions but only by examining the work and living places of their subjects. Aswe can understand people by studying their physical space and belongings, weare now able to investigate users by studying their online activities, postings, andbehavior in a virtual space. This method can be a replacement for traditionaldata collection methods.Among traditional social science data collection techniques, surveys or ex-periments are structured and active, and generating new data is an importantpart of the process. The researcher defines what s/he needs, designs question-naires or experimental treatments, and collects the data based on the results2 Mohammad-Ali Abbasi, Sun-Ki Chai, Huan Liu, Kiran Sagooof administering them. The results provide a greater degree of control over themeasurement but is expensive, time consuming, and may even be dangeroussometimes. On the other hand, studying social media, as an alternative to sur-veys or experiments, can be considered as an extension of passive methods oftraditional social research such as field research and content analysis to observepeople’s attitude.Real-world Behavior Prediction Attitudes1among individuals in a pop-ulation can be determined in social sciences. More specifically, attitudes maybe measured using established data collection techniques, such as surveys, ex-periments, field research, and content analysis. Alternatively, attitudes may bedetermined without direct measurement through models that allow them to bepredicted from individual or collective structural position and/or past actionsand experiences. Such approaches are often described as exogenous and endoge-nous analysis, respectively [5].Online Behavior Prediction The use of World Wide Web content to predictthe attitudes and behavior of individuals or groups is an issue that is increasinglytantalizing and frustrating to social and computer scientists [13]. Conversely,information available online seems to offer a gold mine of useful data - it iscopious, usually publicly accessible, can be located with the aid of search engines,and often has built-in annotations in the form of meta-tags and link information.Furthermore, because such information, when public, can be downloaded byvirtually anyone with an Internet connection, regardless of location, its collectiongenerally incurs less time, expense, intrusiveness, and danger (depending on thepopulation being studied) than traditional primary social science data collectiontechniques such as surveys, experiments, and field research.On the other hand, there is not a straightforward relationship between onlinecontent and attitudes/behavior in the real-world. Although the Web is growingexceedingly fast, some sectors of populations are more likely to use it than others[9]. Furthermore, among the many interesting research issues, novel techniquesare needed to explain the relationship between web content and attitudes ofthose producing the content, and the relationship between these attitudes andactions on the ground by the groups that the content producers represent. Nei-ther relationship is simple to determine, and both require the implementation ofinnovative methodologies in order to provide the impetus to making productiveuse of online information as a predictor of collective behavior.Prediction and understanding of the attitudes and behaviors of individualsand groups based on the sentiment expressed within online virtual communitiesis a natural area of research in the Internet era. Ginsberg et al [6] used Googlesearch engine query data to measure concern about influenza, which in turn wasused to predict influenza epidemics. They used the idea that when many fromspecific area are searching for influenza or topics related to it, this is a signthat there is an epidemic in that place. O’Connor et al. [10] analyzed sentiment1An attitude is a hypothetical construct that represents an individual’s degree of likeor dislike for something.Real-World Behavior Analysis through a Social Media Lens 3polarity of a huge number of tweets and found a correlation of 80% with resultsfrom public opinion polls. Bollen et al [3] used Twitter data to predict trends inthe stock market. They showed that one can predict general stock market trendsfrom the overall mood expressed in a large number of tweets. In other research,Asure and Huberman [2] used