[Paper Review] Effective Semi-Supervised Node Classification on Few-Labeled Graph Data

Ziang Zhou, J. Y. Shi+3 more|arXiv (Cornell University)|Oct 7, 2019

Advanced Graph Neural Networks48 references4 citations

TL;DR

This paper proposes ABN, a framework that significantly improves semi-supervised node classification on graphs with extremely few labeled nodes by using adaptive pseudo labeling, label balancing for skewed data, and a negative sampling regularizer. On Cora with only 1 labeled node per class, ABN boosts GCN accuracy from 44.6% to 62.5% and DAGNN from 59.8% to 66.4%, achieving state-of-the-art performance across four benchmark datasets.

ABSTRACT

Graph neural networks (GNNs) are designed for semi-supervised node classification on graphs where only a small subset of nodes have class labels. However, under extreme cases when very few labels are available (e.g., 1 labeled node per class), GNNs suffer from severe result quality degradation. Several existing studies make an initial effort to ease this situation, but are still far from satisfactory. In this paper, on few-labeled graph data, we propose an effective framework ABN that is readily applicable to both shallow and deep GNN architectures and significantly boosts classification accuracy. In particular, on a benchmark dataset Cora with only 1 labeled node per class, while the classic graph convolutional network (GCN) only has 44.6% accuracy, an immediate instantiation of ABN over GCN achieves 62.5% accuracy; when applied to a deep architecture DAGNN, ABN improves accuracy from 59.8% to 66.4%, which is state of the art. ABN obtains superior performance through three main algorithmic designs. First, it selects high-quality unlabeled nodes via an adaptive pseudo labeling technique, so as to adaptively enhance the training process of GNNs. Second, ABN balances the labels of the selected nodes on real-world skewed graph data by pseudo label balancing. Finally, a negative sampling regularizer is designed for ABN to further utilize the unlabeled nodes. The effectiveness of the three techniques in ABN is well-validated by both theoretical and empirical analysis. Extensive experiments, comparing 12 existing approaches on 4 benchmark datasets, demonstrate that ABN achieves state-of-the-art performance.

Motivation & Objective

• Address the severe performance degradation of graph neural networks (GNNs) when only a handful of nodes are labeled, especially under extreme few-shot settings (e.g., 1 labeled node per class).
• Develop a generalizable framework applicable to both shallow and deep GNN architectures to improve node classification accuracy under low-label regimes.
• Overcome limitations of existing methods that fail to maintain performance quality when label scarcity is extreme.
• Improve model generalization by effectively leveraging unlabeled nodes through intelligent pseudo labeling and regularization techniques.
• Ensure robustness on real-world graphs with skewed class distributions by incorporating pseudo label balancing.

Proposed method

• Introduces an adaptive pseudo labeling technique that dynamically selects high-quality unlabeled nodes based on prediction confidence and graph structure to improve GNN training.
• Applies pseudo label balancing to mitigate the impact of class imbalance in real-world graphs by reweighting or resampling pseudo-labeled nodes during training.
• Designs a negative sampling regularizer that explicitly encourages model discrimination by contrasting positive pseudo-labeled nodes with negative samples from the unlabeled set.
• Integrates all three components—adaptive selection, label balancing, and negative sampling—into a unified framework that enhances GNN performance without architectural changes.
• Employs a confidence-based thresholding mechanism to filter unreliable pseudo labels, ensuring only high-confidence predictions are used in training.
• Supports both GCN and deep architectures like DAGNN, demonstrating broad applicability across GNN types.

Experimental results

Research questions

• RQ1 Can a unified framework significantly improve GNN performance on semi-supervised node classification when only one labeled node per class is available?
• RQ2 How effective is adaptive pseudo labeling in selecting informative unlabeled nodes to boost GNN training under extreme label scarcity?
• RQ3 To what extent does pseudo label balancing improve performance on real-world graphs with skewed class distributions?
• RQ4 What is the contribution of negative sampling regularization in enhancing model generalization on few-labeled graph data?
• RQ5 Does the proposed ABN framework achieve state-of-the-art performance across diverse benchmark datasets under low-label settings?

Key findings

• On the Cora dataset with only 1 labeled node per class, ABN improves GCN accuracy from 44.6% to 62.5%, demonstrating a substantial performance gain over standard GNNs.
• When applied to the deep DAGNN architecture, ABN increases accuracy from 59.8% to 66.4%, setting a new state-of-the-art result on Cora under the same few-shot setting.
• Extensive experiments across four benchmark datasets show ABN consistently outperforms 12 existing methods in semi-supervised node classification under low-label regimes.
• Ablation studies confirm that each component of ABN—adaptive pseudo labeling, label balancing, and negative sampling—contributes significantly to the overall performance gain.
• The framework maintains strong performance across both shallow and deep GNN architectures, proving its generalizability and robustness.
• Theoretical and empirical analyses validate that the three core components of ABN effectively stabilize and improve training dynamics on few-labeled graphs.