End-to-End Influence Maximization in the Field

This paper presents the CHANGE agent for influence maximization, a multiagent problem with many applications in preventative health and other domains. CHANGE addresses major barriers to the deployment of influence maximization by service providers through a series of algorithmic contributions, backed by simulation results on real-world networks. We then conducted a real-world pilot study of CHANGE with a drop-in center serving homeless youth, the first such pilot study of sampling-based influence maximization and only the second study testing any influence maximization agent in the real world. CHANGE obtained comparable influence spread to previously field tested algorithms, but surveyed only 18% of youth to obtain network data. CHANGE has empirical promise in delivering high-quality influence maximization solutions in a manner which can be feasibly implemented by a service provider. While the algorithms underlying CHANGE are easy to implement, they draw on a series of insights into the social behavior of homeless youth. One lesson learned from our study is that, to be successful in the field, algorithms must be designed with their target population and setting in mind. CHANGE both navigates challenges specific to homeless youth (e.g., the difficulty of locating youth to query for edges or serve as peer leaders) and leverages properties of their social network (the friendship paradox). Our experience shows that accounting for both challenges and opportunities in the target population is crucial to produce a practically deployable algorithm.