DeNoise: Learning Robust Graph Representations for Unsupervised Graph-Level Anomaly Detection

Qingfeng Chen, Haojin Zeng, Jingyi Jie, Shichao Zhang, Debo ChengarXiv 2025 · DOI · arXiv

Abstract

Unsupervised graph-level anomaly detection (UGAD) is a critical task in domains such as drug discovery and cybersecurity, where labeled data is scarce. However, traditional UGAD models often fail when training sets are contaminated with unlabeled anomalies. This paper introduces DeNoise, a robust framework designed to learn clean graph representations from noisy data. By utilizing a novel anchor-alignment mechanism and adversarial training, DeNoise effectively suppresses anomalous signals in the latent space. Our evaluation across eight benchmark datasets shows that DeNoise consistently outperforms state-of-the-art methods, particularly as noise levels increase.

Introduction

Graph-level anomaly detection identifies abnormal graphs within a dataset, such as malicious sub-networks or unusual molecular structures. Most existing Unsupervised Graph-Level Anomaly Detection (UGAD) methods operate under the assumption that the training set is entirely composed of "normal" graphs. In practical scenarios, however, training data is frequently contaminated with hidden anomalies. When standard models are trained on such noisy data, they tend to memorize anomalous patterns as part of the normal distribution, leading to a significant degradation in detection performance.

DeNoise addresses this challenge by abandoning the clean-set assumption. It is designed to be robust against contaminated training sets, ensuring that the learned representations remain focused on the dominant normal patterns even when anomalies are present.

Methodology

The DeNoise framework employs a three-step architecture to achieve robustness against noise:

  1. Discriminator & Reconstruction Model: The model uses a Graph Neural Network (GNN) encoder to compress input graphs into latent embeddings. Two decoders then attempt to reconstruct the graph's structure (adjacency matrix) and node attributes. Graphs that deviate significantly from the majority in the latent space are flagged as potential anomalies through a preliminary sorting process.
  2. Encoder Anchor-Alignment Denoising: This core innovation involves extracting highly informative node embeddings from graphs identified as likely normal. Using a Mixup operation, these high-quality "anchor" features are fused into the embeddings of all graphs in the dataset. This process forces anomalous graphs to align more closely with normal patterns in the latent space, effectively "denoising" their representations and preventing the model from learning their unique anomalous features.
  3. Adversarial Alternating Training: The model is trained through a min-max game. It alternates between minimizing reconstruction loss to capture normal patterns and maximizing a contrastive loss that clusters normal graph embeddings while pushing away flagged anomalies.
The three-step architecture of the DeNoise framework.

Finally, anomaly scores are derived by aggregating the mean and standard deviation of both structural and feature reconstruction errors across the multidimensional latent space.

Experimental Setup

DeNoise was evaluated on eight real-world benchmark datasets from the TUDataset collection, including biochemical datasets (COX2, DHFR, AIDS, PROTEINS_full, ENZYMES, PROTEINS, DD) and a social network dataset (IMDB-BINARY). To simulate real-world contamination, the authors injected varying levels of noise into the training sets, with anomaly rates ($\beta$) of 10%, 20%, and 30%.

Results and Discussion

DeNoise demonstrated superior performance and stability compared to existing UGAD baselines across all tested noise levels. While competing models often showed sharp declines in AUC as contamination increased, DeNoise maintained high performance. For instance, on the AIDS dataset with 30% noise, DeNoise maintained a near-perfect 99.87% AUC, while baselines like OCGIN and OCGTL dropped to 84.32% and 89.01%, respectively.

Robustness of DeNoise compared to baselines across varying noise levels on DHFR and ENZYMES datasets.

Key Insights

  • Noise Suppression via Feature Infusion: The fusion of normal node features acts as a regularization technique that "drowns out" anomalous signals, preventing the model from overfitting to noise during the reconstruction phase.
  • Generalization through Diversity: Interestingly, DeNoise's performance improved on certain datasets as noise increased. The successful transformation of anomalies using normal features appears to increase the diversity of the training data, which can enhance the model's overall generalization capabilities.
  • Practical Unsupervised Learning: By proving robustness to contaminated training sets, DeNoise moves the field toward truly unsupervised learning, reducing the reliance on perfectly curated "clean" datasets.

Conclusion

DeNoise provides a powerful solution for graph-level anomaly detection in the presence of noisy training data. Its anchor-alignment and adversarial training mechanisms allow it to learn robust representations that prioritize normal patterns. While the model is sensitive to certain hyperparameters—such as the fusion coefficient ($\lambda$) and the number of anchor nodes ($k$)—it sets a new standard for robustness in UGAD. Future research could focus on developing adaptive methods for automatic hyperparameter tuning to further enhance the model's usability across diverse graph datasets.

Summary updated

Abstract

With the rapid growth of graph-structured data in critical domains, unsupervised graph-level anomaly detection (UGAD) has become a pivotal task. UGAD seeks to identify entire graphs that deviate from normal behavioral patterns. However, most Graph Neural Network (GNN) approaches implicitly assume that the training set is clean, containing only normal graphs, which is rarely true in practice. Even modest contamination by anomalous graphs can distort learned representations and sharply degrade performance. To address this challenge, we propose DeNoise, a robust UGAD framework explicitly designed for contaminated training data. It jointly optimizes a graph-level encoder, an attribute decoder, and a structure decoder via an adversarial objective to learn noise-resistant embeddings. Further, DeNoise introduces an encoder anchor-alignment denoising mechanism that fuses high-information node embeddings from normal graphs into all graph embeddings, improving representation quality while suppressing anomaly interference. A contrastive learning component then compacts normal graph embeddings and repels anomalous ones in the latent space. Extensive experiments on eight real-world datasets demonstrate that DeNoise consistently learns reliable graph-level representations under varying noise intensities and significantly outperforms state-of-the-art UGAD baselines.
BibTeX
@misc{chen2025denoise,
  author = {Qingfeng Chen and Haojin Zeng and Jingyi Jie and Shichao Zhang and Debo Cheng},
  title = {{DeNoise: Learning Robust Graph Representations for Unsupervised Graph-Level Anomaly Detection}},
  year = {2025},
  eprint = {2511.04086},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.04086},
}
Figure