RQPool: A Novel Multi-Branch Graph-Level Anomaly Detection

Abstract

Anomaly Detection (AD) is crucial across various domains,
as it identifies irregularities or unusual patterns
that, if quickly addressed, can prevent financial and data
losses, protect health, and prevent disasters. Many systems
such as social networks, communication systems,
and biological networks are naturally represented as
graphs with entities as nodes and interactions as edges.
By analyzing these structures, we can uncover anomalies
that are not apparent using traditional methods.
However, current Graph-based AD techniques face significant
challenges, particularly low accuracy on larger
datasets. As datasets grow larger, the complexity of
the graphs increases. This complexity makes it more
challenging for models to distinguish normal variations
from true anomalies. Moreover, existing Graph Neural
Network (GNN) algorithms focus primarily on spatial
domain features while neglecting spectral properties.
Furthermore, most algorithms concentrate on intra-graph
properties such as edges and nodes, while overlooking
rich global inter-graph relationships like Graph
Similarity Measures and Cross-Graph Connectivity. To
address these challenges, we propose a novel hybrid
method, RQPool, which integrates intra-graph spectral
properties and inter-graph spatial properties into
a unified Graph-level Anomaly Detection classifier. In
empirical evaluations across multiple datasets, RQPool
consistently achieves higher AUC and macro-F1 scores compared to
purely spectral or spatial baselines, the current state-of-the-
art approaches, particularly excelling on large-scale
graphs.