TAD-Net: Reinforced Anomaly Generation and Wavelet-enhanced Prediction for Temporal Anomaly Detection

ICLR 2026 Conference Submission18679 Authors

Keywords: Anomaly Detection, Graph Data Mining

Abstract:

In dynamic graph environments, structure-based anomaly detection is essential
for applications such as identifying fraudulent calls, fake accounts, and social
bots. While existing methods typically monitor changes in structural features
to detect anomalies, they often fail to account for concept drift—where natu-
ral, gradual changes in network structure are incorrectly flagged as anomalies.
To address this limitation, we introduce Temporal Anomaly Detection NETwork
(TAD-NET), a framework specifically designed to reduce the impact of concept
drift and improve anomalous node detection. TAD-NET consists of three main
components: (i) temporal feature extractor; (ii) reinforced anomaly generator;
and (iii) wavelet-enhanced fusion predictor. The temporal feature extractor identi-
fies changes in node features via dynamic behavior projection, distinguishing be-
tween normal network evolution and true anomalies. Working in tandem with the
anomaly detector, it leverages structural-difference attention to learn robust rep-
resentations for abnormal node detection. To address limited labeled anomalies,
the reinforced anomaly augmenter generates synthetic anomalous samples using
reinforced generative adversarial networks. The wavelet-enhanced fusion pre-
dictor improves adaptability to structural changes by integrating high-frequency
features, maintaining anomaly sensitivity as the network evolves. Experiments
on real-world datasets show that TAD-NET outperforms state-of-the-art methods,
achieving over 6% AUC improvement under concept drift. The code is available
at https://anonymous.4open.science/r/TAD-Net-B26A.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 18679