Pure node selection for imbalanced graph node classification

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Class imbalance, an uneven distribution of data across classes, is a common issue in machine learning, particularly affecting graph-structured data.
• Graph neural networks (GNNs) often assume class balance, leading to performance degradation when faced with imbalanced datasets.
• Existing methods struggle to address both quantity and topological imbalance, and a specific problem termed Randomness Anomalous Connectivity Problem (RACP) arises due to random seed sensitivity in GNNs.

Purpose of the Study:

• To identify and address the Randomness Anomalous Connectivity Problem (RACP) in graph neural networks (GNNs) caused by random seed sensitivity.
• To propose a novel, plug-and-play module that mitigates RACP without requiring specialized algorithms for quantity or topological imbalance.
• To enhance the stability and performance of GNNs on imbalanced graph datasets.

Main Methods:

• Proposed Pure Node Sampling (PNS), a novel plug-and-play module designed for the node synthesis stage.
• PNS operates directly during node synthesis to mitigate RACP and alleviate performance degradation from abnormal neighbor distributions.
• Conducted extensive experiments to analyze the influence of random seeds on GNN performance and validate the effectiveness of PNS.

Main Results:

• Demonstrated that Pure Node Sampling (PNS) effectively eliminates the performance degradation caused by unfavorable random seeds.
• PNS significantly outperforms baseline methods across various benchmark datasets and different GNN backbones.
• Experimental results confirm the effectiveness and stability of PNS in handling class imbalance and RACP in graph data.

Conclusions:

• Pure Node Sampling (PNS) is an effective and stable solution for addressing the Randomness Anomalous Connectivity Problem (RACP) in GNNs.
• PNS offers a versatile, plug-and-play approach to improve GNN performance on imbalanced graph datasets.
• The proposed method enhances GNN robustness against random seed variations and abnormal data distributions.