Abstract
As a popular task in graph learning, node classification seeks to assign labels to nodes, taking into account both their features and connections. However, an important challenge for its application in real-world scenarios is the presence of newly-emerged out-of-distribution samples and noisy samples, which affect the quality and robustness of learned classifiers. Out-of-distribution (OOD) samples are often found in both the training and testing phases. Such samples don’t belong to any known categories. These OOD samples are considered as outliers (OOD noise) when they appear during training, and are recognized as open-set samples during the testing. Meanwhile, in-distribution (IND) noisy data, i.e., known class samples with wrong labels, are also prevalent and inevitably degrade a model’s performance. The challenge of open-set learning with complex IND and OOD noise remains largely unexplored, particularly when dealing with non-IID graph data. To address these challenges, this paper introduces a novel complex-noise-resistant open-set node classification approach, designed for open-set graph data containing both IND and OOD noisy nodes. Specifically, a trustworthiness learner is adopted to learn the trustworthiness rates of the feature and label for each node while a decoder and an open-set classifier are trained to reconstruct the structure of a node and to predict its category simultaneously with the guidance of node trustworthiness. The experimental results demonstrate the superiority of our method.
Introduction
In real-world graph data applications, such as social networks or citation indices, the idealized assumptions of closed-set categories and clean labels often fail. Datasets are frequently plagued by "complex noise," comprising both In-Distribution (IND) noise—where nodes belong to known classes but are mislabeled—and Out-of-Distribution (OOD) noise, where outlier nodes from unknown categories are present during training. Standard Graph Neural Networks (GNNs) tend to be overconfident, forcing these noisy or unknown nodes into known categories, which significantly degrades model reliability. The CONC framework aims to bridge this gap by creating a noise-resistant classifier capable of handling both mislabeled nodes and outliers.
The CONC Framework
To address these challenges, the CONC framework integrates four primary components designed to identify and mitigate noise through adaptive detection.
- GNN Encoder: Aggregates neighborhood information to generate robust node embeddings.
- Trustworthiness Learner: The core innovation that assigns dual trust scores to each node: $r^x$ for feature/structural reliability and $r^y$ for label reliability. This allows the model to distinguish between mislabeled known nodes and complete outliers.
- Adaptive Reconstructor: This module reconstructs node features and degrees. High reconstruction error serves as a signal for OOD samples, as outliers typically deviate from the structural patterns of known classes.
- Robust Open-set Classifier: Utilizes adjusted soft labels. For suspected IND noise, it "sharpens" predictions to favor learned patterns; for suspected OOD noise, it "flattens" predictions to prevent the model from assigning them to known classes.
Key Innovations
CONC moves beyond binary "clean vs. noisy" classifications by recognizing the multidimensional nature of noise. By separating feature trust from label trust, the model achieves more granular data recovery. Furthermore, it leverages the graph's topology (node degrees) as a signal for outlier detection. The use of information-theoretic sharpening and flattening ensures that the classifier remains robust even when training data is heavily corrupted.
Experimental Evaluation
The authors evaluated CONC on citation networks including Cora, Citeseer, Coauthor-CS, and Ogbn-arxiv, injecting varying levels of IND and OOD noise to simulate real-world conditions.
The results demonstrate that CONC significantly outperforms state-of-the-art baselines like PNP and NGC. For instance, on the Ogbn-arxiv dataset with 25% IND noise, CONC achieved an F1-score of 31.00%, compared to 27.67% for the next best method. As shown in the performance curves, CONC maintains stability even under heavy noise (up to 50% corruption), whereas traditional GNN performance typically collapses. In OOD detection tasks, CONC consistently achieved higher AUROC scores, highlighting its effectiveness in distinguishing known from unknown samples.
Conclusion and Future Work
CONC provides a robust solution for node classification in the presence of complex noise. While the current model relies on hyperparameter tuning (specifically $\tau_1, \tau_2$) for noise thresholds, it sets a strong foundation for future work. Potential improvements include automated thresholding mechanisms and extending the framework to dynamic, evolving graphs where new classes emerge over time.
Summary updated