[Paper Review] Mul-GAD: a semi-supervised graph anomaly detection framework via aggregating multi-view information

Zhiyuan Liu, Chunjie Cao+1 more|arXiv (Cornell University)|Dec 11, 2022

Advanced Graph Neural Networks4 citations

TL;DR

Mul-GAD proposes a semi-supervised graph anomaly detection framework that enhances model generalization by fusing multi-view representations through view-level and feature-level fusion strategies. It achieves state-of-the-art performance on multiple benchmark datasets, including PubMed, Amazon Photo, and Books, by leveraging a label-oriented objective function and learnable view weighting to improve detection accuracy and robustness across diverse anomaly types.

ABSTRACT

Anomaly detection is defined as discovering patterns that do not conform to the expected behavior. Previously, anomaly detection was mostly conducted using traditional shallow learning techniques, but with little improvement. As the emergence of graph neural networks (GNN), graph anomaly detection has been greatly developed. However, recent studies have shown that GNN-based methods encounter challenge, in that no graph anomaly detection algorithm can perform generalization on most datasets. To bridge the tap, we propose a multi-view fusion approach for graph anomaly detection (Mul-GAD). The view-level fusion captures the extent of significance between different views, while the feature-level fusion makes full use of complementary information. We theoretically and experimentally elaborate the effectiveness of the fusion strategies. For a more comprehensive conclusion, we further investigate the effect of the objective function and the number of fused views on detection performance. Exploiting these findings, our Mul-GAD is proposed equipped with fusion strategies and the well-performed objective function. Compared with other state-of-the-art detection methods, we achieve a better detection performance and generalization in most scenarios via a series of experiments conducted on Pubmed, Amazon Computer, Amazon Photo, Weibo and Books. Our code is available at https://github.com/liuyishoua/Mul-Graph-Fusion.

Motivation & Objective

• Address the poor generalization performance of existing GNN-based graph anomaly detection methods across diverse datasets.
• Improve detection robustness by fusing multiple graph representations derived from different GNN architectures.
• Investigate the impact of objective functions and number of fused views on model performance to guide optimal design.
• Develop a unified framework that combines view-level and feature-level fusion to enhance representation learning and anomaly detection.
• Establish a theoretically justified fusion mechanism that balances complementary information and suppresses redundancy.

Proposed method

• Generate multi-view node representations using different GNN architectures (e.g., GIN, GAT, GCN, BWGNN) to capture diverse structural and feature-level patterns.
• Implement view-level fusion with learnable attention weights to dynamically assess the contribution of each view to the final anomaly score.
• Apply feature-level fusion by computing a feature similarity matrix to emphasize complementary information and reduce noise from redundant features.
• Utilize a label-oriented objective function in a semi-supervised setting to improve generalization, especially on datasets with limited labeled anomalies.
• Optimize the entire model via stochastic gradient descent using unified representations and the selected objective function.
• Conduct ablation studies on the number of fused views and objective functions to validate design choices.

Experimental results

Research questions

• RQ1 Can multi-view fusion strategies improve the generalization of graph anomaly detection models across diverse datasets?
• RQ2 How does the choice of objective function (label-oriented, reconstruction-oriented, ssl-oriented) affect detection performance and generalization?
• RQ3 What is the optimal number of views to fuse for maximizing detection performance without overfitting?
• RQ4 How do view-level and feature-level fusion mechanisms individually and jointly contribute to improved anomaly detection?
• RQ5 Can learnable fusion weights effectively capture the relative importance of different GNN-based views in anomaly detection?

Key findings

• Mul-GAD achieves the highest AUC of 95.6% on PubMed and 98.7% on Books, outperforming all state-of-the-art methods on these datasets.
• On Amazon Photo, Mul-GAD achieves an AUC of 99.0% (99.2% on validation), significantly outperforming GIN (95.6%) and GAT (98.9%) in the same setting.
• The model demonstrates strong generalization, with AUCs of 91.3% on Amazon Computer and 76.5% on Weibo, outperforming most baselines including OCSVM and LOF.
• Feature-level fusion via similarity matrix computation significantly boosts performance, particularly on synthetic and organic anomalies.
• The label-oriented objective function consistently yields better generalization than reconstruction- or self-supervised alternatives across all datasets.
• Fusing three views (e.g., GAT, GIN, BWGNN) yields optimal performance, while adding more views does not always improve results, indicating a saturation point in fusion benefits.