Efficient parallel computing on the game theory-aware robust influence maximization problem

Abstract

Influence maximization (IM) is a problem of finding the most influential nodes with limited budgets under the given propagation model, which finally maximizes the spread of influence into the network. This work considers the robustness assumption in the IM problem, named the robust influence maximization (RIM) problem, which focuses on the uncertainty factors among the influence propagation models and algorithms, so as to find nodes that can achieve robust performance with different diffusion functions in networks. For modeling the uncertainty, we generate parameter space by adding noises into the independent cascade (IC) propagation model, in which each edge has an interval to deposit its all possible activation probabilities. Then, a novel robust optimal objective function is proposed for the RIM problem, which based on the calculation of the Shapley value, one classical and crucial concept in cooperative game theory. Particularly, to face the computing time explosion for traversing various uncertainty scenarios, we first give a complete parallel computing based framework to improve the proposed algorithm’s efficiency, which will definitely motivate the research that how to fully use the computing resources to speed up the RIM problem solving. Through extensive experiments with nine real-world/synthetic networks and four centrality-based IC models, we find one exciting rule that the worst case on the perturbation interval model occurs when edge probability is equal to the endpoint of the interval, and the influence spread effect is positively correlated with the difference of probabilities between two directions over an edge. The seed nodes selected by the proposed robust algorithm outperforms other state-of-the-art methods under the worst diffusion case.