INFUSE: Influential Node Identification in Multilayer Networks using Multiscale Feature Fusion

Abstract

Multilayer networks offer a powerful framework for modeling complex systems in which nodes engage in multiple types of interactions across interconnected layers. The Identification of influential nodes in multilayer networks is a growing research problem and is relatively less explored. There are numerous studies that have addressed influential node identification in single-layer networks, but very few have focused on multilayer networks. Moreover, existing methods for identifying influential nodes in multilayer complex networks often overlook the complex inter-layer coupling or rely on simple linear heuristics that fail to capture the intricate, non-linear relationship between network topology and spreading dynamics. This paper proposes a model for identifying influence nodes in multilayer networks that leverages multilayer feature fusion, encompassing intra-layer features, weighted centrality features, and inter-layer structural features. The ground truth for learning is obtained through a multilayer Susceptible–Infected–Recovered (SIR) simulation, where infection probabilities are dynamically determined from layer-specific degree distributions, allowing for the realistic modeling of heterogeneous diffusion dynamics. An XGBoost-based ensemble regression model is then trained to learn the non-linear mapping between node-level features and their diffusion-based influence. Furthermore, experiments on nine real-world multilayer datasets demonstrate that the proposed model outperforms classical and heuristic baselines in terms of resilience, generalization, and robustness.