Social Convos: A New Approach to Modeling Information Diffusion in Social Media

A common approach, adopted by most current research, represents users of a social media platform as nodes in a network, connected by various types of links indicating the different kinds of inter-user relationships and interactions. However, social media dynamics and the observed behavioral phenomena do not conform to this user-node-centric view, partly because it ignores the behavioral impact of connected user collectives. GitHub is unique in the social media setting in this respect: it is organized into "repositories", which along with the users who contribute to them, form highly-interactive task-oriented "social collectives". In this paper, we recast our understanding of all social media as a landscape of collectives, or "convos": sets of users connected by a common interest in an (possibly evolving) information artifact, such as a repository in GitHub, a subreddit in Reddit or a group of hashtags in Twitter. We describe a computational approach to classifying convos at different stages of their "lifespan" into distinct collective behavioral classes. We then train a Multi-layer Perceptron (MLP) to learn transition probabilities between behavioral classes to predict, with high-degree of accuracy, future behavior and activity levels of these convos.