Active Learning of Model Parameters for Influence Maximization

Abstract

Previous research efforts on the influence maximization problem assume that the network model parameters are known beforehand. However, this is rarely true in real world networks. This paper deals with the situation when the network information diffusion parameters are unknown. To this end, we firstly examine the parameter sensitivity of a popular diffusion model in influence maximization, i.e., the linear threshold model, to motivate the necessity of learning the unknown model parameters. Experiments show that the influence maximization problem is sensitive to the model parameters under the linear threshold model. In the sequel, we formally define the problem of finding the model parameters for influence maximization as an active learning problem under the linear threshold model. We then propose a weighted sampling algorithm to solve this active learning problem. Extensive experimental evaluations on five popular network datasets demonstrate that the proposed weighted sampling algorithm outperforms pure random sampling in terms of both model accuracy and the proposed objective function.KeywordsInfluence maximizationSocial network analysisActive Learning