GraphST: Class-imbalanced node classification with semantic relation transfer

Jialong Wang

doi:10.1016/j.patcog.2025.112626

GraphST: Class-imbalanced node classification with semantic relation transfer

2025-10-22

0

PRE

AI

DOI：10.1016/j.patcog.2025.112626

Original

Original request for help

Save

Abstract

En 中文

Node classification in Graph Neural Networks often suffers from class-imbalanced issue, which results in model bias towards the majority classes. To alleviate this, recent methods focus on generating positive-incentive noise to augment the representation of minority classes, such as synthesizing new minority nodes or generating pseudo-labeled data. However, these methods fail to capture the relation among classes, which can provide additional knowledge for minority classes augmentation. Recently, semantic information has shown great power in the learning relation among classes, motivating us to use it to solve class-imbalanced node classification. In this paper, we propose a class-imbalanced node classification method GraphST by transferring the class relation from the majority to the minority classes with semantic information and then guiding the node representation learning. Specifically, GraphST has two stages: (1) semantic information generation and adaptation strategy for transferable knowledge learning. (2) Node semantic alignment module for enhancing the positivity of semantic information to facilitate the node classification task. Extensive experiments demonstrate that GraphST performs best on three classic datasets compared with nine baselines.

GraphST: Class-imbalanced node classification with semantic relation transfer

Abstract

Journal

Organization

Cited Papers