Influence Maximization via Hyperbolic Deep Graph Learning in Cyber-Physical-Social Systems

Influence maximization in cyber-physical social systems has an important application background in the field of viral marketing. It attracts extensive research in academic and industrial communities. The state-of-the-art influence maximization algorithms estimate the influence of users on the sampled sub-networks. However, with the explosive growth in data resources of cyber-physical social systems, the generation of these samples becomes expensive. In addition, most existing influence maximization methods rely on the Euclidean space, considering in general only the semantic similarities between users and ignoring the hierarchical relationships between them. In practice though, social networks are hierarchical. Therefore, this study focuses on the influence maximization in the hyperbolic space. Specifically, a hyperbolic influence prediction model is presented, whereby the influence probabilities between users are modeled by incorporating a graph attention network into a hyperbolic GraphSAGE model. Based on influence probabilities, a weighted bipartite network is constructed, and a seed set is found by employing a greedy approach. Finally, numerical experimental studies on real datasets demonstrate that the proposed method not only significantly outperforms state-of-the-art techniques in terms of influence spread and running time, but also outperforms its Euclidean counterpart.