A graph neural network-based node classification model on class-imbalanced graph data

Zhenhua Huang

doi:10.1016/j.knosys.2022.108538

Return

A graph neural network-based node classification model on class-imbalanced graph data

2022-05-01

PRE

Zhenhua Huang *

Yunwen Chen

DOI：10.1016/j.knosys.2022.108538

Original

Save

Abstract

En 中文

Node classification for highly imbalanced graph data is challenging, with existing graph neural networks (GNNs) typically utilizing a balanced class distribution to learn node embeddings on graph data. However, when dealing with an imbalanced class distribution, they tend to bias the nodes of the majority classes while the nodes of the minority classes are under-represented. To overcome this challenge, this work introduces a novel GNN-based Imbalanced Node Classification Model (GNNINCM) that is appropriate for class-imbalanced graph data, comprising two cooperative modules: Embedding Clustering-based Optimization (ECO) and Graph Reconstruction-based Optimization (GRO). ECO first employs a two-layer graph convolutional network (GCN) to obtain node embeddings and then performs clustering analysis to enhance the representative nature of the node embeddings and ease classification. Moreover, GRO employs an inner product decoder to reconstruct graph structure and minimize information loss. In particular, we design a hard sample strategy and integrate it into ECO and GRO to ensure that the embeddings of the hard nodes are correctly represented. Furthermore, we propose a Hard Sample-based Knowledge Distillation Method (HSKDM) to train multiple GNNINCM models simultaneously and improve the overall classification performance. Experiments on three well-known class-imbalanced graph datasets demonstrate that GNN-INCM outperforms current stateof-the-art methods in node classification tasks and that HSKDM can substantially improve the overall classification performance. (c) 2022 Elsevier B.V. All rights reserved.

Keywords:

Graph convolutional network

Class-imbalanced

Node classification

Knowledge distillation

Hard sample