Seeding Influential Nodes in Non-Submodular Models of Information Diffusion

TL;DR

This paper introduces scalable algorithms for selecting influential seed nodes in a complex social network diffusion model that incorporates trust and active participation, improving upon simpler models for better real-world applicability.

Contribution

The paper proposes a projected greedy approach that reduces a complex, non-submodular diffusion model to simpler models, enabling effective seed set selection.

Findings

Algorithms are scalable and effective on synthetic graphs.

The approach performs well on a realistic evacuation network.

Seed sets significantly improve information spread in tested networks.

Abstract

We consider the model of information diffusion in social networks from \cite{Hui2010a} which incorporates trust (weighted links) between actors, and allows actors to actively participate in the spreading process, specifically through the ability to query friends for additional information. This model captures how social agents transmit and act upon information more realistically as compared to the simpler threshold and cascade models. However, it is more difficult to analyze, in particular with respect to seeding strategies. We present efficient, scalable algorithms for determining good seed sets -- initial nodes to inject with the information. Our general approach is to reduce our model to a class of simpler models for which provably good sets can be constructed. By tuning this class of simpler models, we obtain a good seed set for the original more complex model. We call this the \emph{projected greedy approach} because you `project' your model onto a class of simpler models where a greedy seed set selection is near-optimal. We demonstrate the effectiveness of our seeding strategy on synthetic graphs as well as a realistic San Diego evacuation network constructed during the 2007 fires.