Balancing Attraction and Repulsion in Deep Graph Propagation

This direction is merged into another direction: graph over-smoothing (over-smoothing). Its last version stays readable and citable here.

Version 1 of 110 papers18 conceptsfrozen 2026-10-04

Standard Graph Neural Networks (GNNs) are often criticized as being "lazy" learners. As they grow deeper, they tend to act as low-pass filters, smoothing out the differences between nodes until every node representation looks identical. This phenomenon, known as oversmoothing, has long prevented GNNs from reaching the depths common in computer vision or natural language processing.

Recent research has reframed this problem by looking at GNNs through the lens of physical particle systems. By introducing repulsive forces alongside traditional attractive ones, researchers are now building stable, expressive networks that maintain distinct node identities across hundreds of layers.

The Physics of Message Passing

In a typical GNN, message passing is an attractive force. Nodes pull their neighbors toward them in the feature space, which works well for homophilic graphs where neighbors are similar. However, Francesco Di Giovanni and colleagues in Understanding convolution on graphs via energies demonstrate that this is only half the story. They show that graph convolutions can be viewed as physical systems where channel-mixing weights determine whether nodes attract or repel one another.

If a network only uses attractive forces, it inevitably collapses into a single point (oversmoothing). Conversely, if it only uses repulsive forces, it can suffer from "over-sharpening," where features oscillate or diverge. The goal is to find a physical equilibrium.

Yuelin Wang, Shi Jin, and their team took this literally in ACMP: Allen-Cahn Message Passing for Graph Neural Networks with Particle Phase Transition. They treated nodes as particles governed by the Allen-Cahn "double-well" potential—a concept from materials science used to model how different phases of matter separate. By incorporating this potential, the GNN acts like a system undergoing a phase transition, naturally pushing different classes of nodes into separate "wells" or clusters, allowing the model to exceed 100 layers without losing discriminative power.

Structural Balance and Signed Graphs

While the particle analogy provides intuition, it requires a formal framework to ensure the system doesn't explode. This is where signed graphs come in. In a signed graph, edges are either positive (denoting similarity/attraction) or negative (denoting dissimilarity/repulsion).

Jiaqi WANG, Yifei Wang, and James Cheng argue in A Signed Graph Approach to Understanding and Mitigating Oversmoothing that oversmoothing is essentially a "loss of sign." When a network treats all edges as positive, it loses the ability to distinguish between clusters.

They propose Structural Balanced Propagation (SBP). This method organizes interactions based on the principle of structural balance:

  1. Positive edges should connect nodes within the same cluster.
  2. Negative edges should connect nodes in different clusters.

When these forces are balanced, the message-passing dynamics converge to stable, distinct clusters rather than a single mean. This framework allowed the researchers to successfully train GNNs with up to 300 layers.

The Stability Threshold

The primary challenge in using repulsive forces is numerical stability. In particle systems, if the repulsive force is too strong, the "energy" of the features can grow uncontrollably, leading to feature explosion.

As noted by Yixuan Ma and Yu Guang Wang in their work on Hypergraph Message Passing, preventing this explosion requires specific damping mechanisms or forcing terms. There exists a critical "stability threshold":

  • Too little repulsion: The system collapses into a single representation (Oversmoothing).
  • Too much repulsion: Feature values grow toward infinity (Divergence).
  • Structural Balance: The system reaches a topological equilibrium where classes remain distinct and bounded.

Beyond Simple Edges

This balance of forces is also being applied to solve "under-reaching" in sparse graphs. Weigang Lu and colleagues introduced NodeMixup, which strategically mixes features between nodes. To avoid introducing noise, they decouple the graph structure during this process, effectively treating the interpolation as a way to refine decision boundaries without letting conflicting signals propagate through the graph's topology.

By viewing graph learning as a dynamic balancing act between attraction and repulsion, these researchers have moved GNNs beyond the limitations of shallow architectures, enabling deep models that can finally handle the complexity of large-scale, heterophilic data.

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_balancing_attraction_and_repulsion_in_de_13432ec5,
  title = {Balancing Attraction and Repulsion in Deep Graph Propagation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/balancing-attraction-and-repulsion-in-deep-graph-propagation/rdr_13432ec5478441e19c55f8c1ec96bcb8},
  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_balancing_attraction_and_repulsion_in_de_13432ec5_v1,
  title = {Balancing Attraction and Repulsion in Deep Graph Propagation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/balancing-attraction-and-repulsion-in-deep-graph-propagation/rdr_13432ec5478441e19c55f8c1ec96bcb8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}