Abstract

The DARPA SocialSim challenge problem measured participant’s ability, given 30 months of meta-data on user activity on GitHub, to predict the next months’ activity as measured by a broad range of metrics applied to ground truth, using agent-based simulation. The challenge involved making predictions about roughly 3 million individuals performing a combined 30 million actions on 6 million repositories. We describe the agent framework and the models we employed. Our team used a variety of learning methods contributing to six different types of agents that were tested against a wide range of metrics. The broadly most successful method of those tried sampled from a stationary probability distribution of actions and target repositories for each agent. First, we describe the agent-based simulator we developed to carry out massive-scale simulations of techno-social systems. Second, we present the inference methods that we employed to implement different agent-based models, based on statistical modeling of historical activity, graph embedding to infer future interactions, Bayesian models to capture activity processes, and methods to predict the emergence of new users and repositories that did not exist in the historical data. These are novel applications of existing analytical tools to derive agent models from available data. Third, we provide a rigorous evaluation of the performance of six different models, as measured by a wide range of metrics. We also describe

Date

May 8, 2019

Authors

James Blythe, Emilio Ferrara, Di Huang, Kristina Lerman, Goran Muric, Anna Sapienza, Alexey Tregubov, Diogo Pacheco, John Bollenbacher, Alessandro Flammini, Pik-Mai Hui, Filippo Menczer

Conference

AAMAS

Pages

1835-1837