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

TL;DR

This work tackles unsupervised graph-level anomaly detection when training data may be contaminated with anomalies. It introduces DeNoise, a robust UGAD framework that combines adversarial reconstruction, encoder anchor-alignment denoising, and contrastive separation to learn noise-resistant graph embeddings. The model jointly optimizes a graph-level encoder with attribute and structure decoders under a min–max objective and uses a multidimensional reconstruction-score module to produce robust anomaly scores. Across eight real-world datasets and varying contamination levels, DeNoise achieves state-of-the-art performance, demonstrates strong noise resilience, and provides insights into parameter sensitivities and denoising dynamics, indicating practical utility for real-world graph analytics.

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.

Figures (8)

• Figure 1: Illustration of graph-level anomaly detection (GLAD). The labels above the GLAD model denote behavioral patterns learned during training. The $\operatorname{Score}(\cdot)$ function quantifies the deviation of a new sample from these learned patterns, yielding an anomaly score (higher is more abnormal).
• Figure 2: The DeNoise framework comprises three key components: (1) the establishment of a discriminator and a reconstruction model, (2) a noise reduction phase applied to the encoder, and (3) a multidimensional anomaly assessment module. In the model illustration, node color intensity intuitively reflects the amount of information contained in each node: blue represents nodes from normal graphs, red indicates anomalous nodes, and white signifies nodes with low information content. The terms $\operatorname{S_i}$ and $\operatorname{F_i}$ denote the structural and attribute reconstruction errors, respectively, which are aggregated using different functions.
• Figure 3: Performance comparison under different noise levels on (a) DHFR and (b) ENZYMES datasets. The AUC scores (%) of various baseline methods are plotted against increasing levels of injected noise. The proposed method (blue line) consistently outperforms competing approaches and shows strong robustness to noise, while most baselines experience performance degradation as noise levels increase.
• Figure 4: Performance comparison of DeNoise and its variants under different noise conditions. (a) Experiments conducted on clean datasets ($\beta = 0.0$). (b) Experiments conducted under noisy conditions with 30% anomalous samples in the training set ($\beta = 0.3$).
• Figure 5: Hyperparameter analysis of $\lambda$ and $k$ was conducted on eight datasets. Each heatmap displays the AUC score (%) achieved under different $\lambda$ and $k$ combinations, and darker colors indicate better performance.