Influence Maximization With Graph Neural Network in Multi-Feature Social Network

Abstract

The classical influence maximization (IM) prob-lem aims to select a set of nodes as seed nodes for information propagation. Most related studies focus on one-dimensional information dissemination. However, the diffusion of information should be a multi-dimensional (or multi-feature) diffusion process. The user's judgment of information often depends on the combined judgment of multiple-dimensional information. Based on this, we consider a multi-feature diffusion model. In-formation will be decomposed into multi-dimensional features for diffusion. The user will receive these features and integrate them into a complete message. In addition, we consider seed node selection as a network node classification task. Consid-ering that social networks satisfy a graph structure, we use the node classification technique of graph neural networks. This technique classifies the nodes by the network topology to find the set of influential nodes. We propose a multi-feature influence (MFI) algorithm to improve the performance of the prediction. The set of candidate nodes obtained by classification is used as a benchmark to find the set of nodes with maximum influence. Our experiment results on real datasets show that the proposed algorithm has better algorithmic efficiency than the traditional greedy algorithm and better effectiveness than similar heuristic algorithms.