Mining Social Networks for Viral Marketing

Traditionally, social network models have been descriptive, rather than predictive: they are built at a very coarse level, typically with only a few global parameters, and are not useful for making actual predictions of the future behavior of the network. In the past, this was largely due to lack of data: the networks available for experimental study were small and few, and contained only minimal information about each node. Fortunately, the rise of the Internet has changed this dramatically. Massive quantities of data on very large social networks are now available from blogs, knowledge-sharing sites, collaborative filtering systems, online gaming, social networking sites, newsgroups, chat rooms, etc. These networks typically number in the tens of thousands to millions of nodes, and often contain substantial quantities of information at the level of individual nodes, sufficient to build models of those individuals. Assembling these models into models of the larger network they are part of gives us an unprecedented level of detail in social network analysis, with the corresponding potential for new understanding, useful predictions, and their productive use in decision-making. We have begun to build social network models at this scale, using data from the Epinions knowledge-sharing site, the EachMovie collaborative filtering system, and others [1, 6]. Thes