High-Pass Graph Filtering for Heterophilic Networks

87 concepts64 papers1 proposal

Most traditional Graph Neural Networks (GNNs) operate on an intuitive but restrictive assumption: "birds of a feather flock together." In graph theory, this is known as homophily—the tendency for connected nodes to share the same labels or attributes.

To exploit homophily, standard GNN layers perform neighborhood aggregation, which mathematically acts as a low-pass filter. By averaging the features of a node and its neighbors, these models smooth out local variations to extract broad, shared structural patterns.

However, real-world graphs frequently violate this assumption. In credit card fraud networks, fraudsters intentionally connect to benign accounts to camouflage their behavior. In protein-protein interaction networks or chemical molecules, adjacent nodes are often forced to be different. This property is known as heterophily.

When a low-pass GNN encounters a heterophilic graph, its smoothing operation becomes destructive. It washes away the boundaries between distinct classes, causing representations to collapse into an indistinguishable blur—a phenomenon closely tied to over-smoothing. To effectively learn from heterophilic data, a network must do the exact opposite of smoothing: it must perform signal sharpening via high-pass graph filtering.

The Intuition of Signal Sharpening

If a low-pass filter asks, "How am I similar to my neighbors?", a high-pass filter asks, "How am I different from them?"

In a heterophilic environment, the most discriminative information does not reside in the shared commonalities of a neighborhood, but in its sharp transitions and local contrasts. High-pass filtering isolates these high-frequency signals by capturing the gradients between a node and its immediate surroundings.

Low-Pass Filtering (Smoothing):
Node A [0.8]  <--->  Node B [0.2]   ====>   Both become [0.5] (Loss of distinction)

High-Pass Filtering (Sharpening):
Node A [0.8]  <--->  Node B [0.2]   ====>   Contrast amplified (Preserves boundaries)

By focusing on these differences, high-pass filters prevent category confusion. As demonstrated by Luan et al., high-pass filtering mathematically increases the relative distance between different class centers in the feature space. Instead of pulling dissimilar nodes together, it pushes different categories apart, making them far more distinguishable for downstream classifiers.

Implementing High-Pass Filters via Negative Weights

In spatial terms, standard graph convolution aggregates neighbor features using positive weights. To transform this operation into a high-pass filter, models must incorporate negative weights or signed message passing. Subtracting a neighbor's features from a target node's own representation effectively computes a directional derivative, highlighting local discrepancies.

Several architectural frameworks leverage this mechanism:

  • Generalized PageRank (GPR-GNN): Chien et al. showed that by allowing propagation weights to take negative values, a GNN can adaptively switch between low-pass and high-pass filtering depending on the homophily level of the dataset.
  • Signed Propagation: Yang et al. introduced signed weights via hyperbolic tangent activations to capture dissimilarity, while Shi et al. utilized negative edge weights in the graph Laplacian to model repulsive forces between dissimilar nodes.
  • Continuous Kernels: Traditional self-attention mechanisms are bounded by positive weights. To overcome this, Ma et al. parameterized graph convolution kernels as continuous functions capable of learning negative weights, enabling explicit edge detection and community boundary sharpening.

By subtracting lower-order neighbor information from a node's representation—such as through inverse Laplacian propagation (Liao et al.) or negative second-order residual connections (ClenshawGCN)—the network isolates the unique, high-frequency features that make a node stand out from its context.

Multi-Channel and Adaptive Filtering

Real-world graphs are rarely purely homophilic or purely heterophilic; they are often a complex hybrid of both. Consequently, a rigid high-pass filter is just as problematic as a rigid low-pass filter. Modern architectures resolve this by decoupling the signal into multiple frequency bands.

Luan et al. proposed the Adaptive Channel Mixing (ACM) framework, which extracts information through three parallel paths: an aggregation channel (low-pass), a diversification channel (high-pass), and an identity channel. The model adaptively mixes these channels for each node, ensuring that harmful heterophily is counteracted without sacrificing low-frequency structural layouts.

Similarly, Duan developed a Triple Filter Ensemble (TFE-GNN) that explicitly runs low-pass, band-pass, and high-pass filter banks in tandem. In unsupervised settings, Chen et al. designed PolyGCL to contrast low-pass and high-pass spectral views against one another, proving mathematically that incorporating high-pass spectral information minimizes the error bounds of node representations when homophily is low.

The Spectral Perspective

To truly understand why this works, we must look at the graph spectrum. The graph Laplacian matrix possesses eigenvectors and eigenvalues that represent the "frequencies" of the graph. Low eigenvalues correspond to smooth, slowly varying signals (low frequencies), while high eigenvalues correspond to rapidly oscillating signals across connected nodes (high frequencies).

Traditional spectral GNNs focus heavily on the lower end of this spectrum. However, as Li et al. and Ito et al. point out, the structural information that aligns with node labels in heterophilic graphs is distributed across the higher-frequency components of the spectrum.

Graph Laplacian Spectrum:
[ Low Eigenvalues ] -------------------------> [ High Eigenvalues ]
  Low-frequency (Smoothness)                     High-frequency (Contrast)
  Targeted by standard GNNs                      Targeted by High-Pass Filters
  Best for Homophilic Graphs                     Best for Heterophilic Graphs

By designing piece-wise polynomial filters that segment the spectrum (Lingam et al.) or utilizing learnable Chebyshev series to capture the upper end of the Laplacian spectrum (Ito et al.), contemporary architectures can extract micro-level details and local patterns. High-pass graph filtering shifts the paradigm of graph representation learning from simple neighborhood averaging to sophisticated, multi-frequency signal processing.

Key Foundational Work

Diverse Message Passing for Attribute with Heterophily (2021) — contributed 2 concept elements to this direction.
Beyond Low-pass Filtering: Graph Convolutional Networks with Automatic Filtering (2021) — contributed 2 concept elements to this direction.
Revisiting Heterophily For Graph Neural Networks (2022) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_pass_graph_filtering_925a5543,
  title = {high-pass graph filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-pass-graph-filtering/rdr_925a554358184fa1b74f0bc5f45fb871},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_pass_graph_filtering_925a5543_v1,
  title = {high-pass graph filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-pass-graph-filtering/rdr_925a554358184fa1b74f0bc5f45fb871/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake