A Comprehensive Deep Semi-supervised Graph Learning Approach Incorporating Node Re-weighting and Manifold Regularization

Abstract

Deep Graph Neural Networks (GNN) have gained increasing prominence across various fields and applications in recent years. In particular, Graph Convolutional Network (GCN), a specialized GNN, has emerged as a powerful tool for obtaining data representations through the processes of graph-based node smoothing and implementing transformations through neural networks layer by layer. Although GCN excels in the area of semi-supervised learning, it also has its weaknesses. In particular, it 1) tends to overlook the implicit manifold structure of graph data, 2) focuses primarily on graph convolutions, often neglects the graph construction, and 3) rarely addresses topological imbalances. In response to these limitations, we present a comprehensive deep semi-supervised Graph Learning approach incorporating Node Re-weighting and Manifold Regularization (ReNode-GLCNMR). This innovative approach corrects topological inconsistencies in the graph and adaptively reweights labeled nodes. It also seamlessly combines graph learning and graph convolution within one GCN architecture. Moreover, it imposes label smoothing by utilizing an unsupervised loss item. Our experiments on eight benchmark datasets demonstrate that our ReNode-GLCNMR model surpasses the performance of the most advanced semi-supervised GNN methods available.

Jingjun Bi

Former PhD student

I am interested in imbalanced learning, multi-view learning, graph convolutional neural networks, and classification.

Fadi Dornaika

Ikerbasque Research Professor

Ikerbasque Research Professor with expertise in computer vision, machine learning, and pattern recognition.