Overcoming the Homophily Assumption in Graph Neural Networks

2,697 concepts1,314 papers54 proposals

Standard Graph Neural Networks (GNNs) operate on a simple premise: nodes that are connected are likely to be similar. This is the homophily assumption—the network equivalent of the phrase "birds of a feather flock together." For instance, in a citation network, a machine learning paper is highly likely to cite another machine learning paper.

Mathematically, standard message-passing architectures implement this assumption by averaging the features of a node's immediate neighbors. As explained by Harry Zhao et al. in Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks, this neighborhood aggregation acts primarily as a low-pass filter, smoothing out feature variations across the graph topology. While this smoothing works beautifully when the homophily assumption holds, it creates a fundamental vulnerability when applied to graphs where opposites attract.

The Heterophily Challenge: When Smoothing Becomes Destructive

In many real-world networks, connected nodes are intentionally dissimilar. Consider a fraud network where fraudsters connect predominantly to legitimate users, or chemical networks where different types of atoms bind together to form molecules. This property is known as heterophily.

When a standard GNN encounters a heterophilous graph, its low-pass filtering mechanism becomes a liability. By indiscriminately averaging the features of highly diverse neighbors, the model introduces what Yifan Hou et al. describe in Measuring and Improving the Use of Graph Information in Graph Neural Networks as "negative disturbance." Instead of refining a node's representation, message passing washes away its distinct local characteristics.

As Yujun Yan et al. demonstrate in Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks, heterophily and feature over-smoothing are fundamentally interconnected: both stem from an unconstrained drive toward uniformity that destroys the discriminative features needed for classification.

To build models capable of generalizing to heterophilous graphs, researchers have developed three primary mechanisms: high-pass filtering, higher-order neighborhood mixing, and adaptive causal disentanglement.

Mechanism 1: Preserving Details via High-Pass Filtering

To prevent representations from dissolving into a uniform blur, architectures must capture the "missing half" of the graph signal. If standard aggregation is a low-pass filter that captures similarities, the natural countermeasure is to introduce high-pass or band-pass filtering channels that capture differences.

Standard GNN (Low-Pass Only):
Node Features ───► [ Average Aggregation ] ───► Smoothed Representations (Loss of Detail)

Filterbank / Multi-Channel GNN:
               ┌─► [ Low-Pass Filter ] ──┐
Node Features ─┤                         ├─► Adaptive Fusion ──► Distinct Representations
               └─► [ High-Pass Filter ] ─┘

Harry Zhao et al. pioneered this approach by explicitly adding a "diversification" channel alongside traditional aggregation. Instead of just tracking how a node matches its neighbors, the high-pass channel highlights how a node deviates from them.

Similarly, Yimeng Min et al. proposed Geometric Scattering Attention Networks, which adaptively weight low-pass and band-pass information on a per-node basis. By allowing the network to retain high-frequency details, these architectures ensure that nodes remain distinguishable even when surrounded by a neighborhood of a different class.

Further refining this concept, Vijay Lingam et al. developed A Piece-wise Polynomial Filtering Approach for Graph Neural Networks, which segments the graph's frequency spectrum to adaptively extract critical information from connections between dissimilar nodes without manual tuning.

Mechanism 2: Expanding the Horizon to Higher-Order Neighborhoods

Another way to combat local heterophily is to look beyond immediate connections. While a node's immediate 1-hop neighbors might be entirely different from it, its 2-hop or 3-hop neighbors often return to the same class. For example, in a bipartite-like transaction network, a fraudster connects to a victim (1 hop away, completely different), but that victim might connect to another fraudster (2 hops away, highly similar).

[Target Node: Class A] ──(Heterophily)──► [1-Hop Neighbor: Class B] ──(Heterophily)──► [2-Hop Neighbor: Class A]
   │                                                                                             ▲
   └─────────────────────────────────── Higher-Order Receptive Field ────────────────────────────┘

To exploit these multi-hop structural patterns, Sami Abu-El-Haija et al. designed MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing. MixHop allows a single neural network layer to aggregate information from multiple neighborhood distances (0-hop, 1-hop, and 2-hop) simultaneously, learning distinct linear transformations for each distance. This enables the model to understand the relationship between near and far neighbors rather than blending them together.

To manage the noise that often accompanies larger receptive fields, Zhining Liu et al. developed hpGAT: High-Order Proximity Informed Graph Attention Network, integrating matrix factorization directly into the attention mechanism to inject higher-order topological context safely without triggering over-smoothing.

Mechanism 3: Adaptive Trust and Causal Disentanglement

Rather than assuming all edges carry the same type of relationship, modern architectures attempt to dynamically measure how much a node should trust its local neighborhood.

Sitao Luan challenged the traditional view that heterophily is universally harmful in Is Heterophily A Real Nightmare For Graph Neural Networks on Performing Node Classification?. The author demonstrated that if a model can adaptively mix aggregation, diversification, and identity channels, heterophilous structures remain highly informative because the pattern of differences is itself a strong predictive signal.

Taking a causal perspective, Fuli Feng et al. proposed Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method, which uses test-time causal intervention to isolate the true effect of neighbor features versus a node's self-features.

In a similar vein, the authors of Graph Contrastive Learning via Interventional View Generation leveraged causal interventions to disentangle homophilic and heterophilic structures into separate graph views. By routing these distinct views through specialized filters, they prevent representation collapse in unsupervised settings, proving that understanding why an edge exists is just as important as knowing that it does.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 1491 papers
  2. Version 1 · 2026-10-04 · 896 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_homophily_assumption_2b0f33e3,
  title = {homophily assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/homophily-assumption/rdr_2b0f33e3778d480c8aaff2e3cb9a76c0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_homophily_assumption_2b0f33e3_v2,
  title = {homophily assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/homophily-assumption/rdr_2b0f33e3778d480c8aaff2e3cb9a76c0/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake