Contrastive sensitivity learning for unsupervised node importance evaluation in complex networks

Yiming Dai

doi:10.1016/j.knosys.2026.116012

Contrastive sensitivity learning for unsupervised node importance evaluation in complex networks

2026-04-20

0

PRE

AI

DOI：10.1016/j.knosys.2026.116012

Original

Original request for help

Save

Abstract

En 中文

In complex networks, evaluating node importance is fundamental to understanding information diffusion, network robustness, and resource allocation. However, most existing approaches rely on supervised signals or dynamic propagation models, which are often costly to obtain and may not faithfully reflect node importance in complex networks. To address this issue, we propose an unsupervised node importance evaluation framework based on contrastive sensitivity learning. The core idea is to construct attribute-ablated contrastive views by replacing node features with a constant matrix, thereby suppressing attribute heterogeneity and highlighting structural information. Node sensitivity to structural and attribute perturbations is then quantified through embedding discrepancies between the baseline and ablated views, which serve as the basis for importance estimation. Unlike conventional multi-view contrastive learning that enforces representation invariance, our method explicitly amplifies inter-view differences to reveal attribute-dependent significance. The framework is fully unsupervised, compatible with mainstream GNN encoders, and flexible in supporting diverse view-generation strategies. Extensive experiments on public graph datasets (e.g., Cora, CiteSeer, PubMed, and Amazon) demonstrate that the proposed method achieves high rank correlation with traditional centrality measures (Degree, Betweenness, Closeness, and PageRank) and existing GNN-based ranking methods, while consistently outperforming them in SIR diffusion validation. These results confirm that contrastive sensitivity learning not only enhances the accuracy and robustness of node importance evaluation, but also provides a new theoretical perspective for understanding information diffusion and structural vulnerability in complex networks.

Keywords:

node importance

contrastive sensitivity learning

complex networks

unsupervised learning

graph neural networks

Contrastive sensitivity learning for unsupervised node importance evaluation in complex networks

Abstract

Journal

Organization

Cited Papers