Influence Maximization Under Partially Observable Environments

Abstract

The problem of influence maximization is a classic subject to study in the field of network science. It is about finding the top-k important individuals in a network for message dissemination under a particular diffusion model. Each year a number of new research papers are published concerning the same issue. However, most of these methods can only operate in situations where the whole graph is visible to the algorithm which is an unrealistic assumption in many cases. There are many cases where the induced network model of a natural phenomenon is associated with missing links. Discarding these links will lead to serious drawbacks in the result. In this work, we have extended the current state of the art influence maximization algorithms by adding a link prediction heuristic step prior to the actual run of the algorithm. For the purpose of link prediction, we have used exponential random graph models also known as ERGM due to their probabilistic link prediction capabilities. We have shown that this heuristic can significantly improve the effectiveness of influence maximization algorithms and in a diffusion scenario we can have a larger number of infected nodes using the seed nodes of the influence maximization algorithm.