Graph Normalization: Methods & Applications

• Graph Normalization is a family of techniques applied across GNNs, combinatorial optimization, and graph data management to leverage graph structure for scaling and grouping.
• In graph neural networks, normalization dynamically computes statistics from nodes, neighborhoods, or batches to mitigate over-smoothing and accelerate training.
• Extensions include metadata normalization in property graphs that reduce redundant information and enforce schema consistency through dependency analysis.

Graph Normalization (GN) is a polysemous term spanning several research programs rather than a single universally accepted operator. In graph machine learning, it usually denotes normalization layers or normalization operators that act on node, edge, graph, or batch representations in graph neural networks (GNNs), often to improve optimization, preserve expressiveness, or mitigate over-smoothing (Cai et al., 2020, Chen et al., 2020). In combinatorial optimization, it denotes an iterative neighborhood-based normalization dynamics on weighted graphs for differentiable approximation of Maximum Weight Independent Set (MWIS)-type problems (Guigues, 6 May 2026). In graph data management, it denotes schema- and dependency-driven redundancy elimination for labeled property graphs and metadata structures (Schrott et al., 3 Mar 2026, Sa'd et al., 5 Mar 2026). A separate line uses graph propagation to refine normalization statistics even outside native graph inputs, as in arbitrary style transfer (Jung et al., 2020).

1. Terminological scope and conceptual divisions

In the GNN literature, graph normalization is primarily a representation-space operation inserted into message-passing stacks. Typical formulations normalize hidden features after graph aggregation or linear transformation, with statistics computed at node-wise, graph-wise, neighborhood-wise, or batch-wise scope (Cai et al., 2020, Chen et al., 2020). The core design question is not merely whether to standardize activations, but over which graph-defined population the statistics should be computed, and whether centering and scaling should themselves depend on graph structure.

A second meaning is dynamical. Here GN is a graph-structured map that divides each node’s mass by a closed-neighborhood denominator and is iterated until convergence. In this usage, normalization is not a feature-standardization primitive but a local competitive dynamics on graph-supported variables, with direct connections to independent sets, assignment, and relaxed graph optimization (Guigues, 2020, Guigues, 6 May 2026).

A third meaning is database-theoretic. In labeled property graphs (LPGs), graph normalization refers to decomposition under graph-native dependencies, including dependencies within nodes, within edges, and between nodes and edges. The objective is redundancy elimination, dependency preservation, and reduction of update anomalies, rather than optimization of neural training (Schrott et al., 3 Mar 2026). Closely related work extends normalization to the metadata layer itself by externalizing recurring metadata into canonical Trait Nodes connected by HAS_TRAIT relationships (Sa'd et al., 5 Mar 2026).

A recurring source of confusion is the acronym itself. In some non-graph papers, “GN” denotes Group Normalization, not graph normalization; the biomedical U-Net study explicitly uses GN in that sense and not as a graph-native notion (Zhou et al., 2018).

2. Graph normalization as a family of GNN normalization operators

In message-passing GNNs, normalization is typically inserted after the aggregation-plus-linear step. A canonical template used in the literature is

$$H^{(k)} = F^{(k)}\!\left(\mathrm{Norm}\!\left(W^{(k)} H^{(k-1)} Q\right)\right),$$

where $Q$ is the graph aggregation matrix and $F^{(k)}$ is the post-normalization nonlinearity or MLP (Cai et al., 2020). What varies across methods is the scope of the statistics $(\mu,\sigma)$, the treatment of centering, and whether affine parameters are static or graph-conditioned.

A useful taxonomy is given by “Learning Graph Normalization for Graph Neural Networks” (Chen et al., 2020). That work defines four normalization levels over a mini-batch of graphs and then learns a channel-wise convex combination of them.

Scope	Statistics are computed from	Interpretation
Node-wise	One node’s own feature vector	Layer-normalization-like self statistics
Adjacency-wise	All feature elements in a node’s neighborhood	Local structural contrast
Graph-wise	All nodes in one graph	Whole-graph context
Batch-wise	All nodes in all graphs in a batch	Batch normalization across graphs

Its unified GN layer is

$$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$

with nonnegative, channel-wise normalized weights $\lambda_n,\lambda_a,\lambda_g,\lambda_b$ that sum to one (Chen et al., 2020). This formulation treats normalization choice as a learnable design variable rather than a fixed architectural constant.

A particularly influential graph-specific instance is GraphNorm. It performs graph-wise, per-feature normalization across the nodes of each graph, but replaces fixed mean subtraction by a learnable shift: $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$ with

$$\mu_j = \frac{1}{n}\sum_{i=1}^n h_{i,j}, \qquad \hat{\sigma}_j^2 = \frac{1}{n}\sum_{i=1}^n (h_{i,j}-\alpha_j\mu_j)^2.$$

GraphNorm is explicitly presented as an InstanceNorm variant with learnable centering, motivated by the claim that graph-wise normalization accelerates optimization while naive full mean subtraction can damage expressiveness on regular graphs (Cai et al., 2020).

3. Spectral, optimization, and over-smoothing analyses

A major strand of GN research asks not only whether normalization helps empirically, but what geometric or spectral effect it has on deep graph propagation. “Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs” gives a precise characterization for linearized GNNs (Scholkemper et al., 2024). It shows that residual injection preserves components in the Krylov subspace

$\Kr(A, X^{(0)}) = \Span(\{A^{i-1}X^{(0)}_{:,j}\}_{i \in [n], j \in [k]}),$

whereas BatchNorm prevents collapse to a one-dimensional constant subspace through per-column rescaling. Under its assumptions, BatchNorm drives representations to the top-$k$ eigenspace of the centered operator $Q$0, while the centering step itself can distort graph signal by changing the effective operator. This critique motivates GraphNormv2, which replaces fixed centering by a learned projection in a graph-informed spectral subspace (Scholkemper et al., 2024).

A complementary analysis is given through the Graph Neural Tangent Kernel. Under a population degree-corrected stochastic block model, row normalization

$Q$1

is shown to preserve class structure better than symmetric normalization

$Q$2

because the corresponding kernel block structure is less entangled with node-specific degree corrections (Sabanayagam et al., 2022). The same work shows that class information degrades with depth for vanilla propagation, though skip connections retain nonzero separability at infinite depth (Sabanayagam et al., 2022).

PowerEmbed takes a different route by making normalization itself a layer-wise orthogonalization procedure: $Q$3 Under full-rank initialization and an eigengap assumption, the iterates converge to the top-$Q$4 eigenvectors of $Q$5 up to orthogonal transformation (Huang et al., 2022). In that formulation, graph normalization is not a moment-matching device but a subspace power iteration that prevents collapse to a single dominant eigenvector and yields a sequence of local-to-global representations.

These analyses collectively shift the interpretation of GN away from generic “stabilization.” In this literature, normalization may act as a preconditioner, a spectral projector, a rank-preserving rescaling mechanism, or an operator that changes which graph frequencies survive depth.

4. Adaptive, group-aware, and structure-conditioned normalization

A large class of methods departs from fixed statistics by conditioning normalization on learned groups, graph structure, sensitive groups, or subgraph descriptors.

Differentiable Group Normalization (DGN) explicitly targets over-smoothing by learning a soft node partition

$Q$6

forming group-weighted embeddings

$Q$7

normalizing each group independently, and combining them residually: $Q$8 Its motivating diagnostics are the group distance ratio $Q$9 and the instance-information quantity $F^{(k)}$0, both introduced to formalize the claim that useful deep smoothing should compress nodes within groups while separating groups from one another (Zhou et al., 2020).

ResNorm addresses a different graph pathology: long-tailed degree distributions. It uses node-wise mean $F^{(k)}$1, node-wise standard deviation $F^{(k)}$2, a global transformed-NStd mean $F^{(k)}$3, and a node-wise partial shift based on $F^{(k)}$4: $F^{(k)}$5 The scale is intended to reshape the node-wise standard deviation distribution toward a more normal-like form, improving low-degree “tail” nodes, while the shift is designed to avoid the over-smoothing risk attributed to standard mean subtraction (Liang et al., 2022).

FairNorm makes normalization conditional on sensitive groups. For each sensitive subgroup $F^{(k)}$6, it computes group-wise feature statistics

$F^{(k)}$7

and applies

$F^{(k)}$8

It then regularizes normalized group means and deviations to reduce statistical parity and equal opportunity gaps while retaining normalization’s optimization benefits (Kose et al., 2022).

GRANOLA makes affine parameters node- and graph-adaptive through an auxiliary GNN fed with Random Node Features (RNF): $F^{(k)}$9

$(\mu,\sigma)$0

The standardization itself is LayerNorm-node-like, but the affine correction becomes graph-conditioned and node-specific. The paper further proves that removing RNF collapses the method back to standard MPNN-level expressiveness (Eliasof et al., 2024).

SuperNorm embeds a precomputed subgraph-specific factor $(\mu,\sigma)$1 before and after BatchNorm. Its Representation Calibration and Representation Enhancement steps are intended to inject node-induced subgraph structure into normalization, and the paper argues that this pushes arbitrary GNN backbones to be at least as powerful as the 1-WL test while also alleviating over-smoothing (Chen et al., 2023).

R-SoftGraphAIN goes further by normalizing through a graph-wise covariance transform. In its hard form,

$(\mu,\sigma)$2

and its softened practical version uses truncated SVD and interpolation parameters $(\mu,\sigma)$3 to avoid overly rigid whitening (Li et al., 2023). The paper interprets this as anisotropic covariance normalization that preserves global structural knowledge and, when combined with residual and initial connections, avoids both feature forgetting and structural degeneration (Li et al., 2023).

Outside graph-structured inputs proper, graph-conditioned normalization also appears in arbitrary style transfer. Graph Instance Normalization (GrIN) starts from AdaIN and replaces the raw style mean $(\mu,\sigma)$4 with a graph-smoothed mean $(\mu,\sigma)$5 produced by GCN propagation over a mini-batch graph of style-sample similarities: $(\mu,\sigma)$6

$(\mu,\sigma)$7

The graph is dynamic, defined by $(\mu,\sigma)$8, and is used during training only; at test time the graph layers are removed and $(\mu,\sigma)$9 (Jung et al., 2020).

5. Graph normalization as iterative graph dynamics for combinatorial optimization

A separate literature uses “graph normalization” for a local competitive dynamics on graph-supported weights. In its basic form,

$$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$0

or componentwise

$$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$1

where $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$2 is the adjacency matrix and division is well-defined when $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$3 (Guigues, 2020). Iterating this map yields Iterative Graph Normalization (IGN). The paper proves that the binary fixed points are exactly the indicator vectors of maximal independent sets, analyzes their Jacobian stability, and characterizes attraction basins under assumptions on a post-normalization activation $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$4. It also establishes that non-maximal independent sets are repulsive under nonvanishing derivative conditions, but global convergence and binarization are presented there as conjectures rather than general theorems (Guigues, 2020).

A later weighted and regularized formulation makes the optimization connection explicit. With vertex weights $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$5, $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$6, and $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$7, the update becomes

$$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$8

In weighted state space $$\hat{h}_{v_{k,i}}= \gamma \left( \lambda_{n}\odot\hat{h}_{v_{k,i}}^{(n)} +\lambda_{a}\odot\hat{h}_{v_{k,i}}^{(a)} +\lambda_{g}\odot\hat{h}_{v_{k,i}}^{(g)} +\lambda_{b}\odot\hat{h}_{v_{k,i}}^{(b)} \right)+\beta,$$9, the corresponding energy is

$\lambda_n,\lambda_a,\lambda_g,\lambda_b$0

and one has the identity

$\lambda_n,\lambda_a,\lambda_g,\lambda_b$1

The paper derives the update as an exact Majorization-Minimization step, interprets it as a diagonal quasi-Newton descent on $\lambda_n,\lambda_a,\lambda_g,\lambda_b$2, and proves monotone increase of the weighted mass $\lambda_n,\lambda_a,\lambda_g,\lambda_b$3 along the trajectory (Guigues, 6 May 2026).

The same work proves that for $\lambda_n,\lambda_a,\lambda_g,\lambda_b$4, non-binary fixed points are strictly repulsive, binary fixed points correspond to $\lambda_n,\lambda_a,\lambda_g,\lambda_b$5-stable maximal independent sets, and every trajectory from a normalizable initialization converges to a binary MIS while the relaxed MWIS objective strictly increases until convergence (Guigues, 6 May 2026). It also reformulates the dynamics as a nonlinear replicator equation on the simplex, where vertices compete as strategies and average fitness equals the weighted mass. This yields a weighted extension of the Motzkin–Straus correspondence on a weight-tilted simplex and clarifies the connection to assignment normalization: for assignment problems, GN becomes a conflict-graph analogue of Sinkhorn that tends toward hard assignments rather than doubly stochastic matrices (Guigues, 6 May 2026).

6. Graph-native normalization of property graphs and metadata

In knowledge-graph and database research, graph normalization means redundancy removal in labeled property graphs rather than feature normalization in neural networks. “A Graph-Native Approach to Normalization” formalizes an LPG as

$\lambda_n,\lambda_a,\lambda_g,\lambda_b$6

introduces graph object functional dependencies (GNFDs)

$\lambda_n,\lambda_a,\lambda_g,\lambda_b$7

and evaluates them by requiring the relational instance $\lambda_n,\lambda_a,\lambda_g,\lambda_b$8 induced by graph pattern $\lambda_n,\lambda_a,\lambda_g,\lambda_b$9 to satisfy the FD $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$0 (Schrott et al., 3 Mar 2026). GNFDs are classified as within-node, within-edge, or between-graph-object dependencies, and Armstrong-style rules—Reflexivity, Augmentation, Decomposition, Restriction, and Transitivity—are stated for reasoning over them (Schrott et al., 3 Mar 2026).

On that basis, the paper defines graph-native normal forms GN-1NF, GN-2NF, scoped $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$1-GN-3NF and $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$2-GN-BCNF, and global GN-3NF/GN-BCNF (Schrott et al., 3 Mar 2026). Its normalization algorithms construct graph transformations that may move properties, create new nodes, or reify edges so that dependencies within edges and across node-edge boundaries can be normalized without collapsing the LPG model into a purely node-centric representation. Empirically, under the supplied GNFDs, the method reduces measured redundancy metrics to $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$3 in all reported scenarios and scales from toy graphs to million-edge datasets (Schrott et al., 3 Mar 2026).

A related but more metadata-specific line is the Fifth Graph Normal Form (5GNF). Here the central object is the Trait Node, a canonical reusable metadata object connected by HAS_TRAIT edges. The defining rule is that every metadata value with meaning independent of a specific node or relationship type must be represented as a canonical Trait node, and all domain-to-metadata associations must be expressed exclusively through HAS_TRAIT (Sa'd et al., 5 Mar 2026). The framework introduces trait dependencies $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$4, a trait extraction algorithm, and theoretical properties of losslessness, minimality, and uniqueness up to isomorphism for normalized schemas (Sa'd et al., 5 Mar 2026).

Its Northwind evaluation is concrete: before normalization, embedded metadata values are reported as $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$5 with $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$6 distinct metadata values, yielding

$$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$7

After applying 5GNF, the graph contains $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$8 LocationTrait nodes, $$\mathrm{GraphNorm}(h_{i,j}) = \gamma_j \cdot \frac{h_{i,j} - \alpha_j \mu_j}{\hat{\sigma}_j} + \beta_j,$$9 ShippingTrait nodes, $$\mu_j = \frac{1}{n}\sum_{i=1}^n h_{i,j}, \qquad \hat{\sigma}_j^2 = \frac{1}{n}\sum_{i=1}^n (h_{i,j}-\alpha_j\mu_j)^2.$$0 HAS_TRAIT relationships, $$\mu_j = \frac{1}{n}\sum_{i=1}^n h_{i,j}, \qquad \hat{\sigma}_j^2 = \frac{1}{n}\sum_{i=1}^n (h_{i,j}-\alpha_j\mu_j)^2.$$1 remaining embedded metadata, and approximately $$\mu_j = \frac{1}{n}\sum_{i=1}^n h_{i,j}, \qquad \hat{\sigma}_j^2 = \frac{1}{n}\sum_{i=1}^n (h_{i,j}-\alpha_j\mu_j)^2.$$2 redundant metadata values removed, with

$$\mu_j = \frac{1}{n}\sum_{i=1}^n h_{i,j}, \qquad \hat{\sigma}_j^2 = \frac{1}{n}\sum_{i=1}^n (h_{i,j}-\alpha_j\mu_j)^2.$$3

(Sa'd et al., 5 Mar 2026). In this usage, graph normalization is explicitly a schema- and metadata-level decomposition discipline.

Taken together, these lines show that “Graph Normalization” is best understood as a family resemblance term. Across GNNs, differentiable graph optimization, and graph data modeling, the common thread is the use of graph-defined structure to decide what should be rescaled, grouped, centered, externalized, or made mutually exclusive. The objects being normalized, however, range from hidden feature channels and sample graphs, to relaxed combinatorial variables, to schema- and metadata-bearing graph objects.