Detecting anomalies in dynamic graphs—such as identifying a single fraudulent transaction in a global financial network or a lone bot in a social media surge—is a classic "needle in a haystack" problem. The primary challenge is not just the evolving nature of the graph, but the extreme rarity of anomalies. This "sample imbalance" often leads machine learning models to ignore the few anomalies they encounter in favor of the overwhelming majority of normal data.
In their IJCAI-25 paper, "ABNet: Mitigating Sample Imbalance in Anomaly Detection Within Dynamic Graphs," Yifan Hong and colleagues introduce a framework designed to level the playing field for rare events by generating diverse synthetic anomalous samples and highlighting temporal behavioral shifts.
The Challenge of Imbalance in Dynamic Graphs
Most existing anomaly detection methods for dynamic graphs fall into two traps. First, they rely on Graph Neural Networks (GNNs) that aggregate information from neighbors. In an imbalanced dataset, this "smooths out" the unique characteristics of an anomalous node, making it look indistinguishable from its normal neighbors.
Second, because anomalies are so infrequent, models—especially unsupervised or semi-supervised ones—simply do not see enough examples to learn what a "diverse" anomaly looks like. They end up overfitting to the few examples they have or failing to generalize to new types of suspicious behavior as the graph evolves.
The ABNet Framework
ABNet addresses these issues through a three-pronged architecture designed to capture temporal signals and expand the training set with synthetic data.
1. Feature Extractor (Temporal Focus)
Instead of just looking at a node's neighborhood, ABNet uses a projection-based approach. It calculates the "temporal feature difference" between time steps. By isolating the component of a node's features that is orthogonal to its previous state, the model highlights change rather than static similarity, making abrupt behavioral shifts much easier to spot.
2. Anomaly Augmenter (Synthetic Diversity)
To solve the scarcity problem, the authors built an autoencoder that incorporates Discrete Wavelet Transform (DWT). DWT helps the model retain high-frequency, fine-grained details that are often lost in standard compression but are critical for identifying anomalies. By injecting random noise into the latent space during decoding, ABNet generates a wide variety of synthetic anomalous features to train on.
3. Anomaly Detector (Meta-Learning)
The detector uses a dual-channel architecture to capture both deep, abstract patterns and shallow, immediate cues. Crucially, it employs a meta-learning framework (with inner and outer optimization loops). This allows the model to "learn how to learn," enabling it to adapt rapidly to new graph structures and emerging anomaly types in the most recent snapshots.
Experimental Evaluation
The researchers evaluated ABNet on three real-world dynamic graph datasets: Wikipedia (user edit events), Reddit (subreddit activity), and Mooc (student interactions). ABNet consistently outperformed several high-profile baselines (including TGAT, GDN, and SAD) across all datasets using the Area Under the Curve (AUC) metric:
- Wikipedia: Achieved 92.21% AUC, a significant jump over the next best method, SAD (86.71%).
- Reddit: Reached 75.78% AUC, outperforming TADDY and GDN.
- Mooc: Reached 75.34% AUC.
Ablation studies confirmed that the Feature Projection was the most critical component; removing it caused the Wikipedia AUC to plummet from 92.21% to 71.46%, proving that capturing temporal shifts is the "secret sauce" of the framework.
Conclusion
ABNet demonstrates that in dynamic graphs, the anomaly is often found in the deviation from a node's own history, not just its difference from its neighbors. By combining temporal projection with DWT-based augmentation and meta-learning, the framework provides the agility needed to update model parameters quickly as the graph's underlying distribution shifts. Future work will focus on refining augmentation to mimic complex, multi-node attack patterns and deepening the model's understanding of long-term evolutionary dynamics.
Summary updated