High Frequency Graph Neural Networks for Anomaly Detection

This direction is merged into another direction: low-pass filtering bias (low-pass filtering limitation). Its last version stays readable and citable here.

Version 1 of 114 papers17 conceptsfrozen 2026-10-04

Most Graph Neural Networks (GNNs) are built on a fundamental assumption: homophily. This is the idea that connected nodes tend to be similar. To exploit this, standard GNNs act as "low-pass filters," mathematically smoothing the features of a node with those of its neighbors. While this works wonders for classifying typical nodes, it creates a "blind spot" for the very things we often want to find: anomalies.

The Smoothing Problem

In a standard GCN, the aggregation process effectively averages out local differences. As noted in AnomMAN: Detect Anomaly on Multi-view Attributed Networks by Wenhao Yang, this smoothing process inadvertently masks outliers. If an anomalous node is surrounded by benign neighbors, a low-pass filter will force that node to "blend in," erasing the sharp local differences that make it suspicious.

Researchers like Yixin Liu in ARC: A Generalist Graph Anomaly Detector with In-Context Learning argue that this smoothing is fundamentally counterproductive for anomaly detection. If the goal is to find the "odd one out," we shouldn't be using an architecture designed to make everyone look the same.

The Spectral Right-Shift

To understand why anomalies disappear, we have to look at the graph's "spectrum." In graph signal processing, low frequencies correspond to signals that change slowly across the network (smoothness), while high frequencies correspond to signals that change rapidly between neighbors (sharpness).

In Rethinking Graph Neural Networks for Anomaly Detection, Jianheng Tang and colleagues discovered a consistent pattern: the presence of anomalies causes a "right-shift" in the graph's spectral energy. While normal network behavior resides in the low-frequency range, anomalies manifest as high-frequency "vibrations." Because standard GNNs are designed to suppress these high frequencies to reduce noise, they end up throwing away the primary evidence of abnormality.

Capturing the High-Frequency Signal

The emerging solution is to move beyond simple low-pass filters and design GNNs that can "hear" the high-frequency signals of outliers.

One approach, introduced by Ziwei Chai in Can Abnormality be Detected by Graph Neural Networks?, is AMNet. This framework uses learnable multi-frequency filters. Instead of assuming the signal is low-frequency, AMNet adaptively captures signals across the entire spectrum, allowing it to tune into the specific "frequency" of an anomaly.

Another powerful mechanism is the use of wavelets. Unlike standard convolutions that look at a fixed neighborhood, wavelets can act as localized bandpass filters. Jianheng Tang proposed BWGNN, which uses Beta wavelet filters to specifically target the right-shifted spectral energy. This approach has proven versatile; for instance, A Ruhan applied these wavelet-based GNNs to hyperspectral anomaly detection, achieving massive speedups over traditional statistical methods by treating pixels as nodes and catching their high-frequency spectral shifts.

Beyond Simple Outliers

This shift in perspective—from smoothing to frequency analysis—is solving complex problems in diverse fields:

By embracing the high-frequency "noise" that traditional GNNs try to ignore, this research direction has turned a mathematical weakness into a specialized tool for finding the most critical, hidden patterns in complex networks.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_frequency_graph_neural_networks_for_90a994e8,
  title = {High Frequency Graph Neural Networks for Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-frequency-graph-neural-networks-for-anomaly-detection/rdr_90a994e88b9a43fa8bfb91baf9a68459},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_frequency_graph_neural_networks_for_90a994e8_v1,
  title = {High Frequency Graph Neural Networks for Anomaly Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-frequency-graph-neural-networks-for-anomaly-detection/rdr_90a994e88b9a43fa8bfb91baf9a68459/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}