Enhancing Influence Maximization in Social Networks Using Parallel Reverse Reachability Set Computations

Influence Maximization (IM) is the problem of finding a small subset of nodes from a large social network which can potentially spread influence in the maximally. IM finds widespread applications in viral marketing, targeted advertisement, control of epidemics and feed recommendations. In recent years several novel solutions have been proposed [9], [13], [14] for solving IM which have progressively given asymptotically superior results than the previous. In general IM algorithms can take up to several days to find the maximal influence nodes on billion scale social networks. In this work, we carefully observe the execution profile of a state-of-the-art solution (Stop and Stare (SSA) [14]) and investigate opportunities for parallelization. IM algorithms typically involve randomization, and we propose to exploit some of the architectural and programming benefits offered by modern processors so as to achieve quicker execution. We propose a new algorithm for parallel generation and storage of random samples in the SSA algorithm and implement them on both multi-core and many-core processors. We show that our solution provides nearly 1.8x improvement in running times on a large social network, while ensuring that the maximal influence computed using our technique is at par with the original solution.