Hybrid long-range dependency-aware graph convolutional network for node classification

doi:10.1007/s10115-025-02473-8

Return

Hybrid long-range dependency-aware graph convolutional network for node classification

2025-05-28

PRE

DOI：10.1007/s10115-025-02473-8

Original

Original request for help

Save

Abstract

En 中文

Graph neural networks, despite their effectiveness, face inherent limitations such as over-squashing within the message-passing mechanism. This impedes their ability to effectively capture long-range dependencies, thereby constraining their performance on graph mining tasks, such as node classification. Recent studies attempt to address this issue by developing two-step aggregation-based methods, whose key idea is to seek nodes with high similarity in topology or attribute feature space for information aggregation. Despite effectiveness, we observe that previous studies share two common deficiencies: single-type feature-based strategies of sifting nodes, and limited to preserving low-frequency information. These limitations result in the inadvertent inclusion of irrelevant nodes and the neglect of multi-frequency information within long-range dependencies. To this end, we propose a novel method, called hybrid long-range dependency-aware graph convolutional network (HLDGCN) to overcome the above deficiencies. Specifically, HLDGCN first employs a flexible node selection process that considers both topological and attribute features across diverse graph structures. Subsequently, the original graph is transformed into a long-range dependency-aware graph by incorporating additional edges that connect selected nodes. Contrary to prior approaches that focus solely on high-similarity nodes, HLDGCN incorporates both high- and low-similarity nodes, assigning positive and negative edge weights, respectively. This design ensures the preservation of diverse frequency information within the graph. Furthermore, HLDGCN employs a specialized graph convolution layer that utilizes a separated fusion strategy to aggregate information effectively on the transformed graph. Extensive experiments on various benchmark datasets, including homophily and heterophily graphs, demonstrate the superiority of HLDGCN on the node classification task.

Keywords:

Over-squashing

Long-range dependency

Different-frequency information

Node classification