---
title: Normalized Corrected GNN Propagation
url: https://www.emergentmind.com/topics/normalized-corrected-gnn-propagation
type: topic
---

# Normalized Corrected GNN Propagation

Searching arXiv for the cited papers and closely related works.
arxiv_search query="2010.13993 Combining Label Propagation and Simple Models Out-performs Graph Neural Networks"
arxiv_search query="2106.03058 Approximate Graph Propagation"
Normalized corrected GNN propagation denotes a family of graph propagation schemes that preserve degree-aware normalization while adding an explicit correction mechanism to counteract error accumulation, oversmoothing, or both. In the cited literature, the phrase refers to several closely related constructions rather than to a single canonical layer: post-hoc residual diffusion and prediction smoothing in Correct and Smooth, kernelized normalized diffusion with residual or teleportation terms in Approximate Graph Propagation, and an operator-level correction that subtracts the degree-stationary component of the symmetric normalized adjacency in recent spectral theory [2010.13993] [2106.03058] [2606.23572].

## 1. Terminological scope and operator forms

Across these works, “normalized” refers to propagation by degree-adjusted graph operators such as the symmetric normalized adjacency
\[
S = D^{-1/2} A D^{-1/2},
\]
or random-walk normalizations such as \(D^{-1}A\) or \(AD^{-1}\). “Corrected” refers to different but related interventions: propagation of residual errors, addition of a residual or teleportation term, or explicit removal of the stationary spectral component.

| Framing | Normalized operator | Correction mechanism |
|---|---|---|
| Correct and Smooth | \(D^{-1/2}AD^{-1/2}\) or \(D^{-1}A\) | diffuse residuals, then smooth predictions |
| Approximate Graph Propagation | \(D^{-a}AD^{-b}\) | weights \(w_i\), including teleportation or heat-kernel weighting |
| Spectral corrected propagation | \(D^{-1/2}AD^{-1/2}\) | subtract the degree-stationary projector |

This multiplicity is important because it prevents a common conflation. One line of work treats correction as a post-processing stage applied to a predictor’s outputs; another treats correction as a property of the diffusion kernel; a third treats correction as an operator-level spectral subtraction. The shared theme is that plain repeated normalized aggregation is not taken as sufficient.

## 2. Correct and Smooth as post-hoc normalized correction

In "Combining Label Propagation and Simple Models Out-performs Graph Neural Networks" [2010.13993], normalized corrected propagation is implemented by Correct and Smooth (C&S). The setup assumes an undirected graph \(G=(V,E)\), adjacency \(A\), degree matrix \(D\), symmetrically normalized adjacency \(S=D^{-1/2}AD^{-1/2}\), node features \(X\in\mathbb{R}^{n\times p}\), and one-hot labels \(Y\in\mathbb{R}^{n\times c}\). A base predictor \(f\), such as a linear classifier or MLP, is trained only on \(X\) and the training labels, ignoring graph structure in the training objective, and outputs class probabilities \(Z\in\mathbb{R}^{n\times c}\).

The first post-processing stage, **Correct**, propagates residual errors rather than labels. The residual matrix is defined by
\[
E_{L_t}=Z_{L_t}-Y_{L_t},\qquad E_{L_v}=0,\qquad E_U=0.
\]
For the Autoscale variant, the propagated error solves a Zhou-style objective
\[
\hat{E}=\arg\min_{W\in\mathbb{R}^{n\times c}} \operatorname{Tr}\!\left(W^\top (I-S)W\right)+\mu\|W-E\|_F^2,
\]
with iteration
\[
E^{(t+1)}=(1-\alpha)E+\alpha S E^{(t)},\qquad E^{(0)}=E,\qquad \alpha=\frac{1}{1+\mu}.
\]
The paper notes that this diffusion is norm-contractive in the sense that \(\|E^{(t)}\|_2\le \|E\|_2\) for all \(t\). Because the propagated residual may then be too small, Autoscale rescales it using the average labeled \(\ell_1\)-error
\[
\sigma=\frac{1}{|L_t|}\sum_{j\in L_t}\|e_j\|_1,
\]
and sets, for unlabeled nodes,
\[
Z^{(r)}_{i,:}=Z_{i,:}+\sigma \frac{\hat{E}_{i,:}}{\|\hat{E}_{i,:}\|_1}.
\]

The alternative FDiff-scale variant uses random-walk normalization with clamping on labeled nodes:
\[
E_U^{(t+1)}=[D^{-1}AE^{(t)}]_U,\qquad E_L^{(t+1)}=E_L.
\]
This is a harmonic extension in the sense of Gaussian fields and harmonic functions. The paper states that, under this propagation, the maximum and minimum values of entries in \(E^{(t)}\) do not go beyond those in \(E_L\), after which a validation-tuned scalar \(s\) yields
\[
Z^{(r)}=Z+s\hat{E}.
\]

The second stage, **Smooth**, propagates predictions rather than residuals. Training nodes are clamped to their ground-truth labels, while validation and unlabeled nodes are initialized from corrected predictions. The iteration is
\[
G^{(t+1)}=(1-\alpha)G+\alpha S G^{(t)},\qquad G^{(0)}=G,
\]
and the fixed point \(\hat{Y}\) defines the final class probabilities. Conceptually, Correct diffuses model error, whereas Smooth diffuses corrected predictions subject to label anchors.

The procedure is decoupled from representation learning and can be applied to arbitrary base predictors, including GNNs. The paper reports that on ogbn-arxiv, GAT alone achieves \(73.56\%\), GAT + C achieves \(73.86\%\), and SOTA (UniMP) is \(73.79\%\). On US County, GCNII (SOTA) is \(88.08\%\), while GCNII + C is \(89.59\%\). More broadly, the paper states that on ogbn-products, C&S with a linear base model outperforms the OGB leaderboard SOTA (UniMP) with \(+1.53\%\) accuracy, \(137\times\) fewer parameters, and \(>100\times\) faster training time.

## 3. Generalized kernels and scalable approximation

"Approximate Graph Propagation" formalizes a broader class of normalized corrected propagation through the graph propagation equation [2106.03058]
\[
\bm{\pi}=\sum_{i=0}^{\infty} w_i \bigl(D^{-a}AD^{-b}\bigr)^i \bm{x}.
\]
Here \(a,b\in[0,1]\) determine the normalization, \(\{w_i\}_{i\ge 0}\) specify the propagation kernel, \(\bm{x}\) is a graph signal, and \(\bm{\pi}\) is the propagated output. This representation subsumes transition probabilities, Personalized PageRank, heat kernel PageRank, Katz, SGC, GDC, and APPNP.

The normalization regimes are explicit. Setting \(a=b=\tfrac12\) gives symmetric normalization
\[
\hat{A}=D^{-1/2}AD^{-1/2},
\]
the operator used in GCN-, SGC-, APPNP-, and GDC-style propagation. Setting \(a=1,b=0\) gives \(D^{-1}A\), and \(a=0,b=1\) gives \(AD^{-1}\), corresponding to random-walk variants. Setting \(a=b=0\) yields the unnormalized adjacency \(A\), used for Katz.

Within this framework, correction is encoded in the weight sequence \(w_i\). SGC is a pure \(L\)-step normalized diffusion with \(w_L=1\) and all other weights zero. APPNP uses
\[
w_i=\alpha(1-\alpha)^i,
\]
so that
\[
\bm{\pi}=\sum_{i=0}^{\infty}\alpha(1-\alpha)^i \hat{A}^i \bm{x}
= \alpha(I-(1-\alpha)\hat{A})^{-1}\bm{x},
\]
which is a normalized propagation corrected by teleportation. GDC uses the heat-kernel weights
\[
w_i=e^{-t}\frac{t^i}{i!},
\]
which alter the spectral profile by Poisson reweighting of path lengths. Katz uses geometric decay \(w_i=\beta^i\), which controls long walks and ensures convergence.

AGP’s contribution is algorithmic as well as formal. The exact infinite sum is truncated at depth \(L=O(\log(1/\delta))\), and the paper defines an approximate propagation with relative error: for any node \(v\) with \(|\pi(v)|>\delta\),
\[
|\pi(v)-\widehat{\pi}(v)|\le \tfrac{1}{10}\pi(v)
\]
with probability at least \(0.99\). The deterministic algorithm uses sparse power iterations with cost \(O(mL)\) per signal. The randomized algorithm maintains residue and reserve vectors, pushes large contributions exactly, and samples small contributions while preserving unbiasedness. The resulting variance bound leads to expected time
\[
O\left(\frac{L^2}{\delta}\sum_{i=1}^{L}\bigl\|Y_i(D^{-a}AD^{-b})^i\bm{x}\bigr\|_1\right),
\]
and in random-walk-like cases this becomes \(\tilde O(1/\delta)\).

The empirical emphasis is scalability. On Reddit, Yelp, Amazon, and Papers100M, AGP instantiates SGC-AGP, APPNP-AGP, and GDC-AGP. The paper reports about \(10\times\) preprocessing speedup over exact SGC/APPNP/GDC with negligible loss in accuracy. On Papers100M, described as the largest publicly available GNN dataset so far, SGC exact requires about \(20{,}000\) s, whereas SGC-AGP requires less than \(2{,}000\) s with similar accuracy of about \(62\%\). The paper further states that AGP-based models scale to Papers100M with \(1.6\)B edges, while vanilla APPNP and GDC implementations would be infeasible.

## 4. Degree-stationary correction and the oversmoothing problem

"A Spectral Theory of Normalized Corrected GNN Propagation" defines a more specific object: the symmetric normalized adjacency with its degree-stationary component removed [2606.23572]. For an undirected graph with adjacency \(A\), degree matrix \(D\), and no added self-loops in this paper, the standard normalized operator is
\[
\widetilde{A}_{\text{norm}}=D^{-1/2}AD^{-1/2}.
\]
Its top eigenvector is not the uniform vector \(1\), but the degree-weighted vector
\[
v=D^{1/2}1.
\]
The corresponding projector is
\[
\Pi_{\text{stat}}=\frac{vv^\top}{v^\top v}
=\frac{D^{1/2}1\,1^\top D^{1/2}}{1^\top D1}.
\]

The normalized corrected propagation operator is then
\[
\widehat A
= D^{-1/2}AD^{-1/2}
-\frac{1}{1^\top D1}D^{1/2}1\,1^\top D^{1/2}
= \widetilde{A}_{\text{norm}}-\Pi_{\text{stat}}.
\]
A corrected linear GCN backbone propagates by
\[
X^{(k+1)}=\widehat A\,X^{(k)}.
\]

The motivation is oversmoothing. Standard normalized propagation repeatedly amplifies the stationary degree direction. If
\[
\widetilde{A}_{\text{norm}}=\sum_{i=1}^n \lambda_i w_iw_i^\top,\qquad
1=\lambda_1\ge \lambda_2\ge\cdots\ge \lambda_n\ge -1,
\]
then
\[
\widetilde{A}_{\text{norm}}^k x
= \sum_{i=1}^n \lambda_i^k (w_i^\top x)w_i.
\]
When \(|\lambda_i|<1\) for \(i\ge 2\), the signal converges to the one-dimensional subspace spanned by \(w_1=D^{1/2}1/\sqrt{1^\top D1}\). The corrected operator zeroes out that stationary component:
\[
\widehat A v=0.
\]
The paper’s central question is whether, after this correction, repeated propagation still preserves class-discriminative signal at depth \(k=O(\log n)\).

In the binary cSBM analysis, the paper shows that after degree concentration,
\[
\widehat A=\eta ss^\top + R',
\]
where \(s\) is the normalized community label vector, \(\eta=\frac{(p-q)n}{2d}\) is the graph signal strength, and the residual satisfies
\[
\|R'\|\lesssim \sqrt{\frac{\log n}{np}}.
\]
This decomposition captures the intended effect of correction: remove the stationary degree mode while preserving a class-signal direction up to a controlled residual.

## 5. Theoretical guarantees in contextual stochastic block models

The spectral theory develops high-probability recovery guarantees in the Contextual Stochastic Block Model [2606.23572]. In the binary case, nodes belong to two equal communities, intra-class and inter-class edge probabilities are \(p\) and \(q\) with \(p>q>0\), and row features satisfy
\[
x_i\sim \mathcal N(\mu,\sigma^2I_m)\quad\text{or}\quad \mathcal N(\nu,\sigma^2I_m),
\]
with centered means \(\mu+\nu=0\). The normalized graph-signal parameter is
\[
\gamma=\frac{p-q}{p+q}.
\]

The main exact-recovery theorem assumes the dense polylogarithmic regime
\[
p\ge C_d \frac{\log^B n}{n},\qquad B>4,
\]
depth
\[
1\le k\le C_0\log n,
\]
and the graph-signal condition
\[
\gamma \ge C_\gamma k\sqrt{\frac{\log n}{np}}.
\]
Under these conditions, after
\[
X^{(k)}=\widehat A^k X,
\]
the rows of \(X^{(k)}\) are linearly separable with probability at least \(1-n^{-\Omega(1)}\), provided the feature SNR satisfies
\[
\frac{\|\mu-\nu\|}{\sigma}\ge
C_{\rm snr}
\max\left\{
\sqrt{\frac{\log n}{n}},
\left(\frac{C_*}{\gamma}\sqrt{\frac{\log n}{np}}\right)^k \sqrt{\log n}
\right\}.
\]
The paper interprets exact recovery as zero misclassification error up to label symmetry.

The multi-class theorem assumes \(L\) fixed equal-size classes, graph density \(\bar d\ge C\log n\), feature dimension \(m\le n^{C_m}\), and signal dominance
\[
|\lambda|\ge C_0 k(\delta+\epsilon),
\]
where
\[
\lambda=\frac{(p-q)n}{\bar d L},\qquad
\epsilon=C_\epsilon\sqrt{\frac{\log n}{\bar d}}.
\]
With scaled embeddings
\[
X^{(k)}=\lambda^{-k}\widehat A^k X,
\]
the theorem states that, with probability at least \(1-n^{-\Omega(1)}\), at least \(n-n_e\) nodes satisfy
\[
\|x_i^{(k)}-\mu_{\ell(i)}\|<\frac{\Delta}{2},
\]
so they are closer to their true class center than to any other center. The bound on \(n_e\) explicitly separates structural and normalization error from propagated feature noise.

The proof strategy is technically distinctive. The paper reduces the analysis to a one-dimensional centered CSBM using a result from Baranwal et al. (2021), then controls \(\widehat A=\eta ss^\top+R'\) both in spectral norm and entrywise. The key estimate is
\[
\left|e_u^\top (R')^a s\right|
\le \frac{1}{\sqrt n}\left(K\sqrt{\frac{\log n}{np}}\right)^a,
\]
proved through degree clipping, atom expansion, and decorated-walk counting. This entrywise control, rather than spectral control alone, is what supports exact nodewise recovery at logarithmic depth.

## 6. Empirical profile, interpretation, and limitations

The empirical picture across these works is consistent in one narrow sense: corrected normalized propagation is repeatedly presented as a remedy for the failure modes of plain diffusion, but the mechanism differs by context. In C&S, the correction is label-aware and post hoc; in AGP, it is kernel-based and scalable; in the spectral theory, it is operator-level and aimed at the stationary cause of oversmoothing.

A common misconception is that normalized corrected propagation always means adding residual or teleportation connections. The cited literature shows three distinct meanings. C&S propagates residual prediction errors and then smooths predictions [2010.13993]. AGP encodes correction in the weights \(w_i\), so APPNP and GDC are corrected relative to plain powers of \(\hat A\) [2106.03058]. The spectral theory instead subtracts the degree-stationary projector directly from \(D^{-1/2}AD^{-1/2}\) [2606.23572].

Another misconception is that correction eliminates depth-related degradation unconditionally. The theoretical guarantees for \(\widehat A\) hold for \(k=O(\log n)\) under explicit graph-signal and feature-SNR assumptions in the dense polylogarithmic regime. The same paper states that the theorems do not guarantee anything for \(k\) beyond \(O(\log n)\), and it explicitly notes limitations: homophily with \(p>q\), equal-size blocks, analysis of a linear propagation backbone only, and exclusion of very sparse graphs and heavy-tailed degree distributions [2606.23572].

The real-data evaluations in the spectral paper support the oversmoothing interpretation rather than replacing it with a universal superiority claim. On synthetic cSBM graphs, corrected propagation maintains classification accuracy across more layers than standard GCN propagation. On Cora, CiteSeer, PubMed, Reddit, ogbn-arxiv, and ogbn-products, corrected propagation shows consistently smaller depth-dependent degradation. The paper also compares against DropEdge, GraphMamba, and RevGNN, reporting competitive accuracy while keeping the intervention at the operator level [2606.23572].

Taken together, these works establish normalized corrected GNN propagation as a technically specific response to the limits of naive graph diffusion. In one form, it is a post-processing head that propagates residuals and then labels; in another, it is a family of normalized diffusion kernels with restart or spectral reweighting; in a third, it is the degree-stationary-corrected normalized adjacency itself. The literature therefore supports not a single recipe, but a coherent design principle: retain degree-aware normalization, identify the failure mode of plain propagation, and introduce a correction that is explicit in either the propagated signal, the kernel, or the spectrum.

Source: https://www.emergentmind.com/topics/normalized-corrected-gnn-propagation