Fueled by the ubiquity of communications access, networked systems have become pervasive and given rise to behaviors whose evolution depends on both individual decisions and the interactions on the network. Examples of such behaviors include media sharing websites, where user recommendations influence product adoption decisions of other users, or more generally public discussion forums where past voting records of users provide indications on how they may influence each other and, therefore, how initial opinions may determine the outcome of future votes. Understanding the evolution of decisions in such connected settings can, therefore, be of significant social and economic benefit. For example, this can help predict the adoption of new social policies, or more pragmatically the commercial success of a new shared application. The importance of those questions has attracted much recent attention, but due to the complexity of networked interactions, much remains to be done. This project takes a multi-disciplinary approach to tackling these challenging questions, and seeks to build on models from statistical physics developed to capture the interactions of charged particles, which interact with each other in a manner akin to how users influence each other in a social network. If successful, the work can both expand the set of tools available to explore the behavior of networked systems, and offer insight into specific problems of interest.
This project explores two fundamental aspects of networked systems, namely, the formation of opinions in networks, and how adoption decisions are made when they are influenced by network neighbors. Networked systems can be of many different forms, including communication networks, social networks, political networks, geographical networks, etc., and are characterized by the fact that connections between network members influence their interactions. Characterizing these interactions is a complex task. The two main goals of the project are to (i) extend models from statistical physics to apply them to fundamental problems in networked systems; and (ii) empirically validate the predictive abilities of these models. Specifically, the project seeks to leverage and extend the Ising spin glass model, and apply these extensions to problems of opinion formation and adoption decisions in networked systems. Empirical validation of the results will then be sought through comparison to data collected from social media websites.
Introduction The intent of this project was to construct a predictive statistical framework to investigate the mechanisms of opinion formation and evolution in online communities where social interactions take place. Specifically, we are interested in how users’ opinions are influenced by others and what factors may affect the changes in public opinion. This framework allows extrapolating the strengths of communications among individuals in a network. We highlight below the project’s scientific findings and their implications. Understanding social influence in opinion evolution Individuals widely seek others’ opinions before making their own decisions, therefore understanding how users interact online and predicting the outcomes of their complex interactions are of particular importance. Given the widespread adoption of social media applications, users nowadays can easily gain access to information from different resources, read specialized reviews, and check comments posted by other users. Users in such social environment are acting as human filters and/or knowledge contributors, and they are thus "networked" in one way or another as they create, share, and consume via blogs, tweets, discussion boards, and forums. These networks affect the information that individuals access and the opinions they form. This work develops a framework to replicate the process of opinion formation and its subsequent updating. The framework enables us to uncover important, but often hidden, factors which drive user online behaviors. One of these factors is the tie strength between individuals. Some other useful questions can also be answered. For example, if a small number of people hold important positions in a network, do they have more influential power and to what extent? Will the quality of the opinions embedded in conversations matter? Will the sentiment of the shared-mind manifest different patterns? Using real-world data in online forum discussions, our investigation confirms that users do influence each other and consequently affect the formation of public opinion. In addition, our framework, once calibrated with the historical data, can be used to predict the outcomes of future social interactions. We find, for example, that if a pair of users has an average tie strength, a positive post by one user has a 69% chance to induce a positive message from the other, ignoring all other factors. Are opinion leaders more influential? In a social community, a relatively small number of important people, the so-called ‘influentials’, are believed to have a substantial influence on the opinions and decisions of the majority. However, in the Internet era, the critical role of these opinion leaders becomes less obvious when interactions move to the virtual space because (1) the volume of information generated in the Web grows exponentially; and (2) the opinion exchanges and updates are at a much faster pace. We have analyzed users’ online behaviors, and accordingly constructed a measure to identify opinion leaders. When an individual actively participates in various discussions, she is more visible than other users who do not. Her online activities not only indicate her expertise but also give her a greater chance of being recognized as a leader. Our study finds that the opinions expressed by opinion leaders are more persuasive and hence confirms that opinion leaders still play an important role in the online world. Does information quality matter? While the virtual space brings some advantages to users by enabling them to seek information and others’ opinions, there are potential disadvantages that come along. One of them is less control over the quality of information. If an opinion is shared by a marplot that always generates spam messages or by a user who is easily influenced in discussions, its social impact is diminishing. For example, we find that the message of an individual who continuously oscillates between the extremes of opinion becomes less convincing in the discussion over time. Do early or late opinions matter more? An individual’s opinion is formed gradually and affected in a cumulative manner. During a discussion, each user shares his/her opinion and knowledge, and as time passes, the social pressure incurred from conversations could lead the participants to reach a consensus. It could be argued that an opinion expressed earlier would be more influential since it had existed longer. However, we find that the later part of the discussion contains the opinions that build on prior wisdom of many, and thus is more influential. The project does not provide a "final" answer to this multi-faceted question of opinion evolution; it only provides an initial quantitative step towards exploring it. A look at the social interactions in the networked view confirms that the social pressure generated in the system will lead to the convergence of opinions. Clearly, this does not end the exploration of when the agreement will be achieved and under what conditions. But, the quantitative framework offers the tools to obtain insights into factors that could potentially affect the outcomes of interactions.
