[Paper Review] Three Revisits to Node-Level Graph Anomaly Detection: Outliers, Message Passing and Hyperbolic Neural Networks

Jing Gu, Dongmian Zou|arXiv (Cornell University)|Mar 6, 2024

Anomaly Detection Techniques and Applications4 citations

TL;DR

This paper revisits node-level graph anomaly detection by introducing graph-aware outlier injection methods, challenging the necessity of message passing in GNNs, and proposing hyperbolic neural networks with specialized architectures and loss functions. It demonstrates that hyperbolic models, particularly Poincaré-based ones, significantly outperform Euclidean and Lorentz counterparts, achieving up to 43.6% ROC-AUC on Squirrel with path+DICE outliers, while message-passing models underperform due to outlier corruption.

ABSTRACT

Graph anomaly detection plays a vital role for identifying abnormal instances in complex networks. Despite advancements of methodology based on deep learning in recent years, existing benchmarking approaches exhibit limitations that hinder a comprehensive comparison. In this paper, we revisit datasets and approaches for unsupervised node-level graph anomaly detection tasks from three aspects. Firstly, we introduce outlier injection methods that create more diverse and graph-based anomalies in graph datasets. Secondly, we compare methods employing message passing against those without, uncovering the unexpected decline in performance associated with message passing. Thirdly, we explore the use of hyperbolic neural networks, specifying crucial architecture and loss design that contribute to enhanced performance. Through rigorous experiments and evaluations, our study sheds light on general strategies for improving node-level graph anomaly detection methods.

Motivation & Objective

• To address the lack of comprehensive, graph-informed outlier injection methods in benchmarking GAD datasets.
• To investigate whether message passing in GNNs is beneficial for unsupervised node-level anomaly detection.
• To explore the potential of hyperbolic neural networks in improving anomaly detection by leveraging hierarchical graph structure.
• To design specialized architectures and loss functions for hyperbolic GAD that enhance performance.

Proposed method

• Proposes new outlier injection techniques that generate anomalies based on graph structure and context, increasing diversity and realism.
• Employs a comparative evaluation framework contrasting message-passing GNNs with non-message-passing models, including Transformers.
• Introduces hyperbolic neural network architectures using Poincaré and Lorentz models for node representation learning.
• Designs specialized loss functions that enhance separation between normal nodes and outliers in hyperbolic space.
• Applies logarithmic and exponential maps to enable transformer-based attention mechanisms in hyperbolic space.
• Uses reconstruction-based anomaly scoring in autoencoder frameworks, with modifications to handle hyperbolic geometry.

Experimental results

Research questions

• RQ1 How do graph-aware outlier injection methods improve the diversity and realism of anomalies in benchmark datasets?
• RQ2 Does message passing in GNNs consistently improve performance in unsupervised node-level graph anomaly detection?
• RQ3 Can hyperbolic neural networks with proper architectural design outperform Euclidean and Lorentz-based models in GAD?
• RQ4 What role do specialized loss functions play in enhancing anomaly detection in hyperbolic space?
• RQ5 How do attention-based models like Transformers perform in hyperbolic space for GAD tasks?

Key findings

• The Poincaré-based model achieved 43.6% ROC-AUC on the Squirrel dataset with path+DICE outliers, outperforming all baselines.
• Message-passing GNNs showed a performance decline when exposed to injected outliers, suggesting message passing may amplify noise from anomalies.
• The Poincaré model achieved 89.3% ROC-AUC on Amazon with path+DICE outliers, significantly outperforming Euclidean (85.4%) and Lorentz (78.0%) models.
• Transformer-based models in hyperbolic space (Poincaré) achieved 55.5% AP on Amazon with DICE outliers, far exceeding the 38.8% of the Euclidean baseline.
• The Lorentz model underperformed across all datasets, indicating limitations in its optimization and representation capacity for GAD.
• The proposed outlier injection method produced more diverse and graph-informed anomalies, enabling more robust evaluation of GAD methods.