Dual Frequency Spectral Filters for Heterophilic Graph Learning

This direction is split into other directions: dual-pass graph filtering. Its last version stays readable and citable here.

Version 1 of 1110 papers144 conceptsfrozen 2026-10-04

Standard Message Passing Neural Networks (MPNNs) operate on a fundamental assumption: homophily. This is the idea that "birds of a feather flock together"—that connected nodes likely share the same labels or features. In spectral terms, these models act as low-pass filters, smoothing out differences between neighbors to find commonalities.

However, many real-world networks are heterophilic, where "opposites attract." In a fraud network, a fraudster connects to a victim; in a chemical molecule, different atoms bind together. When a low-pass filter is applied to these graphs, it washes away the very differences that define the structure, leading to "over-smoothing" where all node representations become indistinguishable. To solve this, a new direction in graph representation learning integrates high-pass filters to capture the "missing half" of the graph signal.

The Spectral Intuition: Blurring vs. Sharpening

To understand why this works, we can view graph convolution through the lens of signal processing. As Deyu Bo and Xiao Wang explain in Beyond Low-frequency Information in Graph Convolutional Networks, a standard graph convolution is essentially a weighted average of a node and its neighbors.

  • Low-pass filters (Smoothing): These capture the general trend or "consensus" of a neighborhood. They are excellent for homophilic graphs where neighbors are similar.
  • High-pass filters (Sharpening): These capture the differences between a node and its neighbors. By subtracting the local average from the node's own features, the filter highlights how a node stands out from its context.

Research by Sitao Luan and Harry Zhao in Complete the Missing Half argues that high-frequency information is not just noise; it is the critical discriminative signal in heterophilic settings.

Dual-Channel Architectures

The most effective way to handle diverse graph types is to process both signals in parallel. Frameworks like GREET, developed by Yixin Liu and Shirui Pan in Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating, use an edge discriminator to partition the graph.

In this dual-stream approach:

  1. Homophilic edges are passed through a low-pass filter to reinforce similarities.
  2. Heterophilic edges are passed through a high-pass filter to preserve the unique identities of the connected nodes.

A similar logic is applied in the GOAL framework by Yizhen Zheng and Shirui Pan in Finding the Missing-half, which identifies and generates "missing" connections of either type to balance the graph's structural information before filtering.

Beyond Discrete Filters: Continuous and Global Adaptation

While early methods used fixed filters, recent work has moved toward making the frequency response learnable and continuous. Huiyuan Chen and colleagues introduced TEDGCN in From Trainable Negative Depth to Edge Heterophily in Graphs, which treats the "depth" of a GCN as a trainable real number. This allows the model to smoothly transition between smoothing (positive depth) and sharpening (negative depth) based on the data.

Furthermore, researchers have recognized that heterophily often requires looking beyond immediate neighbors. Erlin Pan and Zhao Kang proposed PC-Conv in Unifying Homophily and Heterophily with Two-fold Filtering, which uses Poisson-Charlier polynomials to approximate a global heat kernel. This allows the model to capture long-range structural patterns that local message passing misses, effectively "zooming out" to find similarities in the broader graph topology even when immediate neighbors are different.

Why It Works: The Reaction-Diffusion Analogy

A helpful intuition for this dual-frequency approach comes from physics, specifically the reaction-diffusion equation used by Jeongwhan Choi and Sung‐Bae Cho in RDGCL.

In this model, diffusion acts as a low-pass filter, spreading information across the graph to capture collaborative trends (the "blurring" phase). Simultaneously, a reaction term acts as a high-pass filter, pushing back against the diffusion to preserve the node's original, unique features (the "sharpening" phase). By balancing these two forces, the model ensures that node embeddings are context-aware but not so smoothed that they lose their individual discriminative power.

Papers in this version

Showing 60 of 110 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dual_frequency_spectral_filters_for_hete_669ec47e,
  title = {Dual Frequency Spectral Filters for Heterophilic Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-frequency-spectral-filters-for-heterophilic-graph-learning/rdr_669ec47e9fc34d42a782ce37bfafcfd1},
  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_dual_frequency_spectral_filters_for_hete_669ec47e_v1,
  title = {Dual Frequency Spectral Filters for Heterophilic Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-frequency-spectral-filters-for-heterophilic-graph-learning/rdr_669ec47e9fc34d42a782ce37bfafcfd1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}