Improving Influence Maximization from Samples: An Empirical Analysis

Influence maximization is a crucial problem in the analysis of large and complex networks. The problem asks for a set of nodes of a network from which message dissemination is maximized. This paper focuses on influence maximization from samples (IMFS) where the network is hidden. Instead, the input is a set of input-output samples of the influence function over a network. The recently proposed algorithm by Balkanski, Rubinstein, and Singer (BRS) for monotone submodular optimization achieves theoretically-tight optimization ratio. Empirically, however, no work is done to evaluate the actual performance of the BRS algorithm against parameters such as sample size, sample set size, and network topology. This paper provides an empirical analysis of algorithms for IMFS. In particular, we extends BRS by (1) factoring in effects caused by the complex network topology when approximating the marginal contribution of nodes and (2) removing constraints on the sampling distribution. We conduct experiments using two information diffusion models, and samples generated on both synthetic random networks and real-world networks. In general, our proposed algorithm significantly improves the output quality from the BRS algorithm.