Geometry Before Features: Using Ricci Curvature to Weight Graph Neighbors

Introduction

Most of the progress in graph neural networks over the past several years has revolved around one question: how much should a node listen to each of its neighbors? The answer almost always involves features. Graph Attention Networks compute attention scores from node embeddings; GraphSAGE samples and pools feature vectors; even the humble Graph Convolutional Network weights by node degree, which is a coarse proxy for structure. What unites these approaches is that they treat the graph topology as a fixed scaffold and let learned or heuristic feature-based weights do the discriminative work.

The paper Curvature Graph Neural Network (CGNN) by Li, Cao, Zhu, Liu, Zhu, and Wu argues that this is a missed opportunity. The geometry of the graph itself, independent of any node feature, already tells you a great deal about which edges carry trustworthy signal and which edges are likely to corrupt it. Their tool for extracting this information is Ollivier's Ricci curvature, a discrete analog of the curvature that measures how a curve deviates from a straight line in Euclidean space. In this post I want to unpack why curvature is a sensible structural prior, how CGNN turns it into aggregation weights, and where I think the idea has both real value and real limits.

Why Curvature Encodes Structural Importance

The core intuition is elegant. Ollivier's Ricci curvature is defined on an edge by comparing the neighborhoods of its two endpoints using optimal transport. Concretely, you place a probability measure over the neighbors of each endpoint and compute the Wasserstein distance between these two measures relative to the graph distance between the nodes. The result has a clean interpretation:

• Positive curvature means the two endpoints share many common neighbors. The edge sits inside a tightly connected cluster. Think of an edge between two members of the same community who also know all the same people.
• Zero curvature corresponds to a locally flat structure, like an infinite grid graph where every node is structurally equivalent.
• Negative curvature means the endpoints have few common neighbors. The edge behaves like a bridge connecting two otherwise separate groups.

This maps directly onto a problem that has dogged GNNs: oversmoothing. Under the homophily assumption, GNNs work by making neighboring representations more similar. That is useful when neighbors share a class label, but harmful when an edge connects nodes of different classes. Those cross-class edges dilute the useful signal of a target node, and repeated aggregation eventually makes everything look the same.

Here is the connection the paper makes precise. Bridge edges, the ones connecting different communities, tend to be exactly the cross-class edges that drive oversmoothing. And bridge edges are precisely the negatively curved ones. So curvature gives you a way to identify the dangerous edges using topology alone, before training begins and before you look at a single feature vector.

The paper illustrates this with a clean example. Consider a node a whose neighbors are mostly blue (same class) plus one node b that bridges to a different group. The standard methods fail to downweight b properly:

• Equal weighting (Neural FPs, GraphSAGE, GIN) treats b identically to the blue neighbors.
• Degree-based weighting (GCN, SGC, TAGCN) actually does the wrong thing, because b has a low degree and GCN assigns more weight to low-degree neighbors.
• Attention (GAT) cannot know b is harmful until after training, since its weights are data-driven and emerge only through optimization.

Curvature handles this directly. In the worked example the edge (a, b) has a Ricci curvature of -0.50, much lower than the curvature between a and its blue neighbors. The geometry flags b as a bridge immediately.

From Curvature to Aggregation Weights

A curvature value is not a weight, and turning one into the other is where the engineering happens. CGNN introduces two modules to make curvature usable as aggregation weights across heterogeneous datasets.

First, a Negative Curvature Processing Module. Ricci curvature is unbounded below in principle and can take negative values, but aggregation weights generally need to be non-negative and well-behaved. This module transforms the raw curvature so that the negative-to-positive ordering is preserved while producing usable weights.

Second, a Curvature Normalization Module. Different datasets have very different curvature distributions; a social network and a citation graph do not share the same scale of structural overlap. Normalization makes the curvature signal comparable and adaptive across datasets rather than baking in a fixed scale.

This design is a deliberate contrast with the earlier CurvGN model, which also used Ricci curvature but fed it through a multilayer perceptron to produce weights. As the authors note, prior work argued that CurvGN's gains came mostly from the learning capacity of the MLP rather than from the curvature itself. CGNN's response is to make the transformation from curvature to weights more explicit and less reliant on a learned black box, so that any performance gain can be attributed to the structural signal rather than to extra parameters.

On the empirical side, the paper reports two kinds of evidence. On synthetic datasets, where the topology can be controlled, CGNN shows that exploiting curvature significantly improves performance, which supports the claim that the model is genuinely using structure. On real-world data, CGNN outperforms baselines on five dense node-classification benchmarks. The emphasis on dense graphs is worth noting and I will return to it.

My Take: A Useful Prior, Not a Universal One

I find the central claim compelling: structure alone can predict which neighbors dilute your signal. This is a genuinely different stance from the feature-centric mainstream, and it has a practical advantage. Curvature is computed once, before training, from the adjacency structure. It is a fixed, interpretable prior that does not consume model capacity and does not need gradients to discover. In a field where attention mechanisms add parameters and training instability, getting discriminative power "for free" from geometry is attractive.

That said, I want to be honest about the boundaries of the idea.

The first limitation is the homophily assumption. Curvature downweights bridge edges because, under homophily, bridges connect different classes. But on heterophilous graphs, where useful information often comes precisely from structurally distant or cross-community neighbors, this prior could actively hurt. The same negatively curved bridge that CGNN suppresses might be the edge a heterophily-aware model wants to amplify. The five benchmarks reported are the usual homophilous citation-style datasets, so the method's behavior on heterophilous graphs remains an open question.

The second limitation is computational cost. Ollivier's Ricci curvature requires solving an optimal transport problem per edge. On dense or large graphs this is expensive, and I suspect the focus on dense benchmarks is partly a positioning choice: dense graphs are where structural overlap is rich enough for curvature to discriminate well, but they are also where computing it is most costly. Scaling curvature computation, or approximating it cheaply, is a practical bottleneck worth flagging.

The third point is more conceptual. Curvature and features are not really competitors; they are complementary signals. Structure tells you which edges are bridges; features tell you whether a particular bridge happens to be informative anyway. A model that combines a curvature prior with feature-based attention, letting geometry set a baseline that learned weights can override when the data warrants, seems more robust than committing fully to either. CGNN leans hard on the structural side to make its point cleanly, but the production-grade version of this idea is probably a hybrid.

Conclusion

Curvature Graph Neural Network makes a clean and underappreciated argument: the geometry of a graph encodes which neighbors matter, and Ricci curvature extracts that information directly. By downweighting negatively curved bridge edges that drive oversmoothing and trusting positively curved intra-cluster edges, CGNN brings a topology-first perspective to a problem that has been dominated by feature-first methods.

The open questions I find most interesting are about generalization. Does the curvature prior survive on heterophilous graphs, or does it need to be inverted? Can cheaper curvature approximations preserve the discriminative signal at scale? And most broadly, what is the right way to fuse a fixed geometric prior with adaptive feature-based attention? The paper's lasting contribution may be less the specific architecture than the reminder that graph theory has decades of structural tools waiting to be connected to neural networks.