NSF Award
2414308
Social media has had a profound impact on various aspects of human lives, influencing how people communicate, access information, and interact with the world. Their impact is mainly delivered via the vehicle of social influence diffusion, the spreading of information cascades through the network. Example applications of social influence diffusion includes deploying advertising campaigns with the goal of maximizing brand awareness or having youth generate excitement around posts encouraging the prevention of obesity. From a computational perspective, the core of influence diffusion tasks is a decision-making problem that aims at computing an effective decision in response to a query, where the underlying system is often not completely known. This project seeks to develop novel methods to understand influence diffusion. In line with our research agenda, educational efforts will be devoted to curriculum design and encouraging the participation of underrepresented groups as well as K-12 students.<br/><br/><br/>The overall goal of this project is to establish query-decision regression as a principled decision-making diagram to understand social influence diffusion. In direct contrast with the conventional learn-and-optimize pipeline that computes decisions based on models assembled with separately learned components, we neither assume either a priori knowledge of the diffusion rules nor a parametric family of the energy function via deep architectures. Instead, we explore the principle of learn-for-optimization, seeking to push the learning process toward generating direct predictions that can maximize the aimed quantities. The proposed framework is statistically principled, with its foundations built on function approximation, hypothesis realizability, and generalization bounds. Our algorithmic development targets several key components, including loss-augmented inference based on the submodular nature of the diffusion process, cyclic training schemes that can seamlessly connect the discrete and continuous modules, and simulation mechanisms for generating task-aware estimators. The theoretical parts will be supplemented by empirical studies over real-world datasets of social influence cascades, covering a great variety of management tasks such as targeted viral marketing, information source detection, outbreak detection, and active friending.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
