Standard Graph Neural Networks (GNNs) operate on a foundational assumption: "birds of a feather flock together." In graph machine learning, this is known as homophily—the tendency for connected nodes to share similar attributes or labels. Mathematically, traditional graph convolutions act as low-pass filters, averaging features across neighborhoods to smooth out differences.
However, real-world networks are rarely purely homophilic. Many graphs exhibit heterophily, where connected nodes are intentionally dissimilar—such as fraud networks where fraudsters connect to victims, or purchase graphs connecting buyers to products. When a traditional low-pass GNN encounters a heterophilic structure, it forces dissimilar neighbors to look alike, washing away the critical variations needed for classification or clustering. As shown by Deyu Bo et al. in Beyond Low-frequency Information in Graph Convolutional Networks, relying exclusively on low-frequency signals fundamentally limits a model's ability to adapt to diverse, disassortative network topologies.
To break this "homophily trap," a prominent research direction has emerged: dual-frequency spectral filtering. Instead of forcing a graph into a single filtering regime, these methods process graph signals through parallel low-pass and high-pass channels, capturing both local smoothness and sharp feature variations simultaneously.
The Mechanics of Dual-Frequency Filtering
To understand why dual-frequency filtering works, it helps to view graph convolutions through the lens of spectral graph theory. A graph signal can be decomposed into different frequency components using the graph Laplacian:
- Low-frequency signals correspond to smooth transitions across the graph, where connected nodes have similar values. Low-pass filters preserve these signals, making them ideal for extracting shared semantics and homophilic context.
- High-frequency signals correspond to sharp transitions, where connected nodes differ significantly. High-pass filters amplify these differences, making them ideal for isolating distinctive node features and heterophilic boundaries.
By combining both signals, a network can adapt to any topology. For example, You Li et al. designed a multi-task autoencoder framework in Graph Representation Learning Beyond Node and Homophily that explicitly models node pairs using both similarity-based (low-frequency) and difference-based (high-frequency) signals. This dual-signal capture ensures that the model remains expressive regardless of whether connected nodes are similar or distinct.
Disentangling Homophilic and Heterophilic Views
A common strategy for implementing dual-frequency filtering is to explicitly separate or "disentangle" the graph's underlying structure into distinct homophilic and heterophilic views before applying targeted filters.
┌───► Homophilic View ───► Low-Pass Filter ───┐
│ ▼
Raw Graph Structure┤ Combination ──► Node Embeddings
│ ▲
└───► Heterophilic View ──► High-Pass Filter ──┘
This structural separation can be achieved in several ways:
- Edge Discrimination: In Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating, Yixin Liu et al. introduce a framework that automatically distinguishes between homophilic and heterophilic edges. It applies low-pass filters to smooth features across like-to-like connections and high-pass filters to sharpen differences across heterophilic connections, preventing over-smoothing.
- Causal Intervention: Rather than relying on heuristics, Zengyi Wo et al. propose Graph Contrastive Learning via Interventional View Generation. This method uses causal intervention to split the graph into separate homophilic and heterophilic views, routing them to specialized low-pass and high-pass spectral filters to preserve local distinctions during contrastive learning.
- Relational and Multimodal Separation: In heterogeneous and multimodal settings, the need for separation is even more pronounced. Yaming Yang et al. leverage this concept in Enhancing Homophily-Heterophily Separation: Relation-Aware Learning in Heterogeneous Graphs to isolate relation-specific signals. Similarly, in Disentangling Homophily and Heterophily in Multimodal Graph Clustering, Zhaochen Guo et al. demonstrate that applying low-pass filters to homophilic structures captures cross-modal consensus, while high-pass filters on heterophilic structures amplify unique, modality-specific variations.
Dual-Pass and Multi-Filter Architectures
Instead of explicitly separating edges beforehand, another branch of work builds dual-frequency capabilities directly into the neural architecture using parallel branches or polynomial filter approximations.
In Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication, Maysam Behmanesh et al. propose a dual-pass encoder featuring two complementary Graph Convolutional Network (GCN) branches. One branch acts as a low-pass filter to capture broad structural context, while the parallel branch operates as a high-pass filter to emphasize node-neighbor differences. Concatenating the outputs of both passes yields highly discriminative node embeddings.
Real-world graphs, however, are rarely uniformly homophilic or heterophilic; they are usually a complex, hybrid mixture of both. To address this, researchers have turned to adaptive polynomial filters that can approximate arbitrary spectral responses:
- Two-Fold Filtering via Polynomials: Bingheng Li et al. introduce PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering, which uses Poisson-Charlier (PC) polynomials to approximate a global heterophilic heat kernel. This allows a single architecture to adaptively balance local feature smoothing and global structural extraction. The authors later simplified and robustified this approach in Simplified PCNet with Robustness by utilizing a continuous, learnable filter order that integrates local and global graph information seamlessly.
- Filter Ensembles: Rather than relying on a single polynomial family, Rui Duan et al. propose Unifying Homophily and Heterophily for Spectral Graph Neural Networks via Triple Filter Ensembles. By adaptively weighting low-pass and high-pass filter ensembles, the model dynamically tunes its frequency response based on the specific homophily-heterophily ratio of the input graph.
Generalization Across Graph Tasks
By moving beyond simple low-pass smoothing, dual-frequency architectures achieve strong generalization across a wide variety of downstream graph learning paradigms:
- Graph Clustering: Traditional clustering algorithms fail when similar nodes are not directly connected. In Beyond Homophily: Reconstructing Structure for Graph-agnostic Clustering, Erlin Pan et al. show that mixed filtering captures both smooth similarities and sharp transitions, allowing clustering algorithms to succeed without prior knowledge of the graph's homophily level.
- Graph Complementary Learning: Real-world graphs often suffer from missing edges. He Zhang et al. develop the GOAL framework in Finding the Missing-half: Graph Complementary Learning for Homophily-prone and Heterophily-prone Graphs, which infers missing connections and applies a dual-frequency convolution to handle both homophily-prone and heterophily-prone missing structures.
- Contrastive Learning and Domain Adaptation: In S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning, Guancheng Wan et al. combine dual-frequency spectral filters with a fast MLP encoder to maintain high accuracy on heterophilic graphs while significantly accelerating inference. Furthermore, when transferring knowledge between different graphs, Ruiyi Fang et al. demonstrate in Homophily Enhanced Graph Domain Adaptation that mixed spectral filters can separately align homophilic and heterophilic signals, ensuring robust domain adaptation even when the "social rules" of the graph change between the source and target domains.
Key Foundational Work
Finding the Missing-half: Graph Complementary Learning for Homophily-prone and Heterophily-prone Graphs (2023) — contributed 2 concept elements to this direction.
Beyond Low-frequency Information in Graph Convolutional Networks (2021) — contributed 2 concept elements to this direction.
Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating (2023) — contributed 1 concept elements to this direction.