A universal sampling method based on feature and structural comprehensive proximity measure

Jinhui Pang

doi:10.1016/j.neucom.2025.131189

A universal sampling method based on feature and structural comprehensive proximity measure

2025-08-05

0

PRE

AI

DOI：10.1016/j.neucom.2025.131189

Original

Original request for help

Save

Abstract

En 中文

• We propose a novel universal Feature-Structure Sampling (FSS) method based on the comprehensive proximity measure, which is plug-and-play compatible with existing GNN models, enabling accelerated computation and enhanced performance. • We utilize a graph attention network to extract feature information and first introduce Hodge score, grounded in the Hodge decomposition theorem to capture structural information. Furthermore, we theoretically demonstrate that Hodge score effectively calculates the global significance of each node within the graph structure. • To accelerate training and reduce the impact of heterogeneous edges, FSS samples the top k important nodes based on comprehensive proximity. This selection captures the majority of the neighborhood information while effectively filtering out heterogeneous data. • Experimental results show that FSS not only accelerates training speed but also mitigates the impact of heterogeneous edges, thereby enhancing model performance.

Keywords:

Feature-Structure Sampling

Graph Neural Networks

Hodge Score

Comprehensive Proximity

Node Importance

A universal sampling method based on feature and structural comprehensive proximity measure

Abstract

Journal

Organization