Title: A Lightweight Deep Learning-based Model for Ranking Influential Nodes in Complex Networks

Abstract: The identification of influential nodes within complex networks is a vital objective with broad implications across numerous fields. Nevertheless, current methodologies frequently struggle to balance the competing demands of precision and computational speed. To overcome these limitations, this study introduces 1D-CGS, a streamlined and robust hybrid framework that merges the rapid processing capabilities of one-dimensional convolutional neural networks (1D-CNNs) with the topological representation strengths of GraphSAGE to facilitate efficient node ranking. The architecture relies on a concise input representation derived from two fundamental topological metrics: node degree and the average degree of neighboring nodes. These metrics undergo 1D convolution to capture local structural patterns, after which GraphSAGE layers are employed to aggregate information from the surrounding neighborhood. We define the node ranking challenge as a regression task, utilizing the Susceptible-Infected-Recovered (SIR) model to establish ground truth influence scores. The 1D-CGS model undergoes initial training on synthetic datasets generated via the Barabasi-Albert model before being deployed on real-world networks to pinpoint influential nodes. Comprehensive evaluations across twelve real-world networks reveal that 1D-CGS markedly surpasses both conventional centrality metrics and contemporary deep learning approaches in ranking precision, all while maintaining exceptionally low runtime costs. Compared to the top-performing deep learning baselines, the model delivers an average enhancement of 4.73% in Kendall's Tau correlation and 7.67% in Jaccard Similarity. Additionally, it attains an average Monotonicity Index (MI) of 0.99, yielding nearly ideal rank distributions that signify highly distinct and discriminative rankings. Moreover, all experimental results validate that 1D-CGS executes within a highly efficient timeframe, operating significantly faster than existing deep learning techniques, thereby rendering it well-suited for large-scale implementation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC