Introduction
Graph Neural Networks (GNNs) typically assume that training labels are accurate and that the set of classes remains constant between training and testing. However, real-world graph data is often "messy," containing In-Distribution (IND) noise (mislabeled nodes), Out-of-Distribution (OOD) noise (outlier nodes), and open-set scenarios where entirely new categories appear during testing. The CONC framework, presented at IJCAI-24, addresses these challenges simultaneously by learning to detect and adapt to noise at the node level.
The CONC Architecture
CONC moves away from rigid data filtering toward a "soft" adaptive learning paradigm. The architecture consists of three primary components: a Trustworthiness Learner, an Adaptive Reconstructor, and a Robust Open-set Classifier.
Trustworthiness Learner
This module estimates two distinct scores for each node: feature trustworthiness (detecting outliers) and label trustworthiness (identifying mislabels). These scores allow the model to weight the influence of each node during training rather than discarding potentially useful data, which is critical in graph settings where node connectivity is vital.
Adaptive Reconstructor
A graph auto-encoder is employed to reconstruct both node features and node degrees. Since noisy or outlier nodes often exhibit structural anomalies, high reconstruction errors serve as a strong signal for noise detection. The reconstructor uses feature trustworthiness scores to prioritize learning from clean structural patterns.
Robust Open-set Classifier
The classifier utilizes adjusted soft labels to mitigate noise impact. For suspected IND noise, it "sharpens" predictions to favor learned patterns over faulty labels. For suspected OOD noise, it "flattens" predictions toward a uniform distribution, preventing the model from over-confidently misclassifying outliers into known classes.
Experimental Results
The researchers evaluated CONC on standard citation networks, including Citeseer and Coauthor-CS, under various noise conditions.
As shown in the results, CONC demonstrates significant resilience to heavy noise. Even at a 50% label noise rate, the model maintains stable performance, outperforming state-of-the-art robust GNNs. On average, CONC achieved a 4.37% improvement in F1-score and a 3.85% improvement in AUROC across all benchmarks. It also proved significantly better at identifying "unknown" classes during testing, reducing the rate of incorrect forced classifications.
Conclusion
CONC provides a robust solution for node classification in the presence of complex noise. By integrating trustworthiness estimation with structural reconstruction and adaptive label adjustment, it effectively bridges the gap between theoretical GNN performance and real-world data challenges. Future work includes extending this framework to dynamic graphs and more diverse feature noise types.
Summary updated