GADY: Unsupervised Anomaly Detection on Dynamic Graphs

Yuyang Tian

Keywords: dynamic graph, anomaly detection

Abstract:

Anomaly detection on dynamic graphs aims to
identify entities that exhibit abnormal behaviours compared to
the standard patterns observed in the graphs and their temporal
information. It has attracted increasing attention due to its appli-
cations in various domains, such as finance, network security, and
social networks. However, existing methods face two significant
challenges: (1) dynamic structure capture challenge: how to
capture graph structure with complex temporal information
effectively, and (2) negative sampling challenge: how to construct
high-quality negative samples for unsupervised learning. To
address these challenges, we propose a Generative Anomaly
Detection on Dynamic Graphs (GADY). GADY is a continuous
dynamic graph model that can capture fine-grained temporal
information to tackle the dynamic structure capture challenge,
overcoming the limitations of existing discrete methods. Specifi-
cally, we propose to use Prioritization Temporal Aggregation and
Status Features to boost the dynamic graph encoder for anomaly
detection. For the second challenge, we introduce a novel use of
Generative Adversarial Networks to generate negative subgraphs.
In addition, auxiliary loss functions were introduced in the
generator training objective, ensuring the generated samples’
diversity and quality simultaneously. Extensive experiments show
that our proposed GADY significantly outperforms the state-
of-the-art method on three real-world datasets. Supplementary
experiments further validate the effectiveness of our model design
and the necessity of each component.

Submission Number: 10