Influence Maximization in Multi-layer Social Networks Based on Differentiated Graph Embeddings

TL;DR

This paper introduces Inf-MDE, a novel multi-layer influence maximization method that uses differentiated graph embeddings and adaptive influence aggregation to better identify influential nodes in complex social networks.

Contribution

It proposes a new multi-layer influence maximization approach with differentiated graph embedding and adaptive message passing, addressing limitations of existing GNN-based methods.

Findings

Inf-MDE outperforms state-of-the-art methods on four datasets.

The adaptive local influence aggregation improves influence modeling.

Model effectively captures inter-layer and intra-layer influence dynamics.

Abstract

Identifying influential nodes is crucial in social network analysis. Existing methods often neglect local opinion leader tendencies, resulting in overlapping influence ranges for seed nodes. Furthermore, approaches based on vanilla graph neural networks (GNNs) struggle to effectively aggregate influence characteristics during message passing, particularly with varying influence intensities. Current techniques also fail to adequately address the multi-layer nature of social networks and node heterogeneity. To address these issues, this paper proposes Inf-MDE, a novel multi-layer influence maximization method leveraging differentiated graph embedding. Inf-MDE models social relationships using a multi-layer network structure. The model extracts a self-influence propagation subgraph to eliminate the representation bias between node embeddings and propagation dynamics. Additionally, Inf-MDE incorporates an adaptive local influence aggregation mechanism within its GNN design. This mechanism dynamically adjusts influence feature aggregation during message passing based on local context and influence intensity, enabling it to effectively capture both inter-layer propagation heterogeneity and intra-layer diffusion dynamics. Extensive experiments across four distinct multi-layer social network datasets demonstrate that Inf-MDE significantly outperforms state-of-the-art methods.