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

TL;DR

DeNoise is a novel framework for unsupervised graph-level anomaly detection that effectively handles contaminated training data by learning noise-resistant embeddings through adversarial, denoising, and contrastive mechanisms, outperforming existing methods.

Contribution

DeNoise introduces a robust UGAD framework with an adversarial training scheme, encoder anchor-alignment denoising, and contrastive learning to improve anomaly detection in noisy training environments.

Findings

Consistently outperforms state-of-the-art UGAD methods.

Effectively handles varying noise levels in training data.

Demonstrates strong generalization across eight real-world datasets.

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.