Graph neural networks (GNNs) classify a node by repeatedly mixing its features with those of its neighbours. That mixing can be read in two ways. In the spatial view, the model looks at particular edges and nodes. In the spectral view, the model keeps some "graph frequencies" of the feature signal and suppresses others. Low frequencies are patterns that vary slowly across connected nodes, such as a community's shared topic. High frequencies are patterns that flip sign across edges, such as the boundary between two groups.
A whole line of architectures now adjusts this frequency response node by node, so that core nodes are smoothed and boundary nodes are sharpened. Every explanation tool we have for GNNs, however, still answers only the spatial question: which edges mattered. Nobody can ask a trained model, for a single node, "did you classify this from its community-level signal or from how it differs from its neighbours?" and get an answer that has been checked.
We propose node-level spectral attribution. For each node, we remove one frequency band of the input signal at a time and measure how much the model's prediction for that node changes. This produces a per-node "frequency profile" of what the prediction depended on. We then validate the profile on synthetic graphs where the correct band is planted by construction. Once validated, it serves as a single instrument for three questions that are currently argued with indirect evidence. First, do the routing decisions of adaptive spectral models actually reflect what their predictions use? Second, does local heterophily (neighbours tending to have different labels) predict which bands a node relies on? Third, does depth-induced collapse show up as a shift of reliance toward the lowest frequencies?
Core Claim
Band-removal attribution in the graph Fourier domain is a faithful, node-level explanation, and it can be checked against planted ground truth. Our main bet is that it disagrees with two things the field currently treats as explanations. The first is the gate weights of mixture-of-experts spectral models. The second is a node's local homophily ratio. Where it disagrees, it is the one that predicts what happens when the band is actually removed.
The closest prior idea is Spectral Occlusion. It occludes Fourier bands of images to explain CNNs and reports an average sensitivity gain of 0.38 over spatial baselines. Images live on a regular grid with a fast Fourier transform and have no notion of node-varying frequency needs. The graph setting differs on three counts:
- The Fourier basis is the eigenbasis of the graph Laplacian, and it is too expensive to compute on large graphs.
- The interesting object is a per-node profile, not one global heatmap.
- There are architectures whose internal routing claims to be a per-node frequency choice, and these can be audited.
The Attribution and Its Construction
Let $L$ be the normalized graph Laplacian. Let $X$ be the node feature matrix and $f_v(X)$ the model's logit for node $v$'s predicted class. Partition the spectrum $[0,2]$ into $B$ bands, and let $P_b$ be the operator that keeps only band $b$ of a graph signal. On graphs small enough for full eigendecomposition, $P_b = U_b U_b^\top$, where $U_b$ are the Laplacian eigenvectors with eigenvalues in band $b$. The attribution of band $b$ for node $v$ is
$$ a_v(b) = f_v(X) - f_v\big(X - P_b X + P_b \tilde X\big), $$
where $\tilde X$ is a reference signal whose band-$b$ content replaces the original. We use two references and report both: zero, and the band content of $X$ with node rows randomly permuted. Removing a whole band can push inputs off the training distribution, and the permuted reference keeps the band's energy while destroying its alignment with the graph.
On graphs of 10,000 to 50,000 nodes, such as the clean heterophily suite, we replace $U_b U_b^\top$ with a Chebyshev polynomial approximation of the band indicator. This avoids the cubic cost of eigendecomposition. One ablation measures how much the approximation distorts attributions on graphs where both are computable.
From $a_v$ we derive a normalized profile per node and two summary quantities:
- the low-frequency share, which is the fraction of attribution mass in the lowest band;
- the profile entropy, which is how spread the node's reliance is across bands.
Attributions are computed on input features, so the method is architecture-agnostic. It applies equally to a spatial GCN, to GAT with ego-neighbour separation, and to adaptive spectral models.
Research Program
Ground truth. We first build synthetic graphs where the correct frequency profile is known per node. The approach extends contextual stochastic block models with region-specific homophily. Within dense homophilic blocks, labels are a function of the low-frequency component of a planted feature signal. In planted heterophilic boundary regions, labels are a function of a high-frequency component. In a third region, labels depend on a mid band, motivated by the observation of Zhang et al. that real graphs often need non-monotone, multi-band responses. Distractor signal is injected into the bands that do not determine the label. The ground truth is then which band determines each node's label. Existing explanation benchmarks such as BA-shapes, Tree-grid (Evaluating Attribution for Graph Neural Networks), SynAnchor and SynMotif (SEMGNN) plant spatial motifs, not frequency dependence. We keep them only as a check that spectral attribution does not degrade on motif tasks when the planted motif has a clear spectral signature.
Auditing adaptive spectral models. We use the validated attribution to audit NODE-MOE. NODE-MOE routes each node to low-pass, high-pass, or constant ChebNetII experts. A GIN gate (Graph Isomorphism Network) makes the routing decision from the local discrepancy $|AX - X|$, where $A$ is the normalized adjacency. A filter-smoothness loss accompanies the gate. Its gate weights are naturally read as "this node needed a high-pass filter," but no one has checked that reading. We define gate faithfulness as the rank correlation, across nodes, between the gate's high-pass weight and the node's measured high-band share $\sum_{b \in \text{high}} a_v(b) / \sum_b |a_v(b)|$.
The motivating concern comes from two findings:
- GAT attention coefficients correlate poorly with true attribution (Evaluating Attribution for Graph Neural Networks).
- Popular explainers can lose to random edges on faithfulness (Probing GNN Explainers).
Gates are a learned soft-routing signal of the same kind as attention, and we expect a similar gap.
We run the same audit on two global spectral models: SAF (Rethinking Spectral GNNs with Spatially Adaptive Filtering) and its BernNet base. SAF shows that non-negative Bernstein filtering implicitly creates signed non-local edges, yet its learned filter is shared by all nodes. Node-level attribution tests whether its predictions are nonetheless node-varying in their frequency reliance, which its global filter alone cannot reveal.
Heterophily and homophily metrics. We test whether the spectral profile carries information beyond local homophily. Two Sides of the Same Coin sorts nodes into regimes by node homophily and relative degree. The global-metric critique in the heterophily literature suggests these coarse statistics miss class-conditional structure. We regress each node's band profile on local homophily, relative degree, and $|AX - X|$, and report the residual variance. We then ask whether that residual predicts something useful, namely the node's error rate when the model's dominant band is removed.
Depth and collapse. We use attribution to address the open dispute about what causes depth-induced collapse:
- Filterbank work such as Complete the Missing Half and EXPASS (Explanation Directed Message Passing) treats it as an aggregation, low-pass phenomenon.
- Interpreting Deep GCNs with Spectrum Perspective blames weight matrices.
- The Oversmoothing Fallacy blames activation plus transformation, and trains residual, batch-normalized GCNs to 1,024 layers.
These accounts make different predictions about input-frequency reliance. If aggregation drives collapse, the low-frequency share of attribution should rise with depth in vanilla GCN before accuracy falls. It should also rise in residual, batch-normalized GCN wherever that model degrades at all. If transformation and activation drive collapse, vanilla GCN's attribution should become diffuse or noisy, with high entropy and no low-band concentration, as the weights become ill-conditioned. Residual, batch-normalized GCN's profile should then stay stable with depth.
We track these quantities alongside effective rank (Erank), which Are we measuring oversmoothing correctly? showed correlates with accuracy at an average Pearson $r = 0.91$ while Dirichlet energy does not. The spectral profile is not meant to replace Erank as a collapse detector. It is meant to say what kind of information is lost, which a rank number cannot.
Baselines
The naive reference point for attribution is a random band profile: bands are ranked uniformly at random for each node. This plays the role of the Random Edge baseline that Probing GNN Explainers found often beats GNNExplainer and GraphMASK on unfaithfulness. Spectral attribution must clearly beat it on deletion faithfulness, or the project stops. A second naive reference is a node-agnostic global profile, the same frequency reliance for every node. We obtain it by averaging attribution over the graph, or by reading off the learned filter response of BernNet or SAF. If per-node profiles are no more faithful than this global one, node-level spectral explanation is unnecessary.
The strongest existing competitor is Integrated Gradients (IG). Evaluating Attribution for Graph Neural Networks found it the most consistent attribution method after CAM (class activation maps), which applies only to specific architectures. That paper's headline result is qualitative; the file gives no numeric score to beat. We run IG two ways:
- in its usual spatial form, with feature attributions projected onto Laplacian bands afterwards;
- directly with respect to band coefficients, which we call spectral IG.
Both use the same trained models, the same band partition, and the same compute budget for path integration as our occlusion passes. On a fully linear filter, occlusion and gradient attributions coincide, so any gap must come from nonlinearity. We will report honestly if spectral IG matches occlusion everywhere. In that case the contribution becomes the benchmark and the audits rather than the estimator, and we would recommend the cheaper method.
For the gate audit, the baselines are NODE-MOE's own gate weights and local node homophily used directly as the explanation. Both are scored against ground-truth bands on synthetic graphs and against deletion faithfulness on real ones. Beating them means two things. On synthetic graphs, our attribution must recover the planted band with higher AUROC than either baseline. On real graphs, removing the band our method ranks first must change predictions more than removing the band the gate or homophily ranks first.
The key ablation compares exact eigenprojection with polynomial-approximated bands, and zero with permuted references. This isolates whether conclusions depend on the approximation needed for scale. For predictive models, NODE-MOE reports 62.31% accuracy on Squirrel versus 57.12% for a single expert. Squirrel is duplicate-contaminated, so we retrain all models on filtered Chameleon and Squirrel and on the clean suite, with matched hyperparameter search budgets. We do not quote those numbers as targets.
Why not just…?
Why not just read NODE-MOE's gate weights as the explanation? The gate is cheap and already per-node. However, it is a learned soft-routing signal of the same kind as attention, and attention fails attribution checks (Evaluating Attribution for Graph Neural Networks). The gate sees only $|AX - X|$, while downstream experts and the shared readout can compensate for a mis-routed node. The gate-faithfulness correlation settles the question directly. If it is high and gates match planted bands as well as occlusion does, we will say gates are adequate explanations for NODE-MOE. Our method would then remain necessary only for models without gates.
Why not just use local homophily as the per-node frequency explanation? This is the most serious threat. The standard story is that heterophilous nodes need high-pass filtering, and if a node's homophily ratio predicts its band profile, explanation adds nothing. We measure the residual of the profile after regressing on homophily, degree, and feature discrepancy. If the residual is small and carries no predictive value for band-removal error, we will report that homophily suffices. That would be a useful negative result for the adaptive-filtering literature. Our expectation is that class-conditional structure, such as two-hop label return of the kind exploited by MixHop, creates mid-band reliance that homophily cannot express. The mixed-band synthetic region is designed to expose exactly this case.
Why not just add residual connections and normalization and stop worrying about frequencies? The Oversmoothing Fallacy and A critical look at the evaluation of GNNs under heterophily suggest simple residual and ego-neighbour designs match or beat specialized spectral models. In the latter, such models took 13 of the top 15 spots on the clean benchmarks. We accept this as a constraint on architecture claims, and we do not propose a new predictor. The question here is explanatory: what a well-trained residual, ego-separated model actually relies on. Spectral attribution applies to those strong simple models just as well, so the strongest predictors are among the things we explain.
Why not just use effective rank to study depth? Erank detects that collapse happens, with its 0.91 correlation to accuracy. It cannot distinguish low-frequency smoothing from noisy, ill-conditioned representations, because both lower rank in the same way. The depth experiment exists to make that distinction, and we report Erank alongside the profile at every depth.
Proposed Experiments and Decisive Checks
The models are GCN, GAT with ego-neighbour separation, GGCN, BernNet, SAF, NODE-MOE, and residual, batch-normalized GCN at 2 to 64 layers. They are trained on Cora, Citeseer, PubMed, filtered Chameleon and Squirrel, Roman-empire, Amazon-ratings, Minesweeper, Tolokers, Questions, and the planted-band synthetic graphs. The metrics are:
- AUROC of recovering the planted band on synthetic graphs;
- deletion and insertion faithfulness, measured as the prediction change from removing or keeping the top-ranked band against random-band removal;
- stability of profiles under small feature perturbations, following the stability criteria of Probing GNN Explainers;
- gate-faithfulness correlation;
- residual variance beyond homophily;
- low-frequency share and profile entropy as functions of depth, reported alongside Erank.
Four outcomes are decisive:
- Spectral attribution must recover planted bands better than spatial IG, spectral IG, gate weights, and homophily, and must beat random bands on deletion faithfulness on every real benchmark. Failing this ends the project.
- The gate audit must give a clear answer, high or low correlation, that holds on the clean heterophily suite and not only on contaminated benchmarks.
- The depth experiment must separate the two collapse accounts. A rise in low-frequency share before accuracy drops in vanilla GCN, absent in the residual, batch-normalized variant, supports aggregation-driven collapse in plain models. Diffuse, high-entropy profiles that track the weight-conditioning story support the transformation account.
- The polynomial-approximation ablation must change band rankings on fewer than a small minority of nodes. Otherwise, conclusions on the 50,000-node graphs cannot be trusted.
References
- Spectral Occlusion - Attribution Beyond Spatial Relevance Heatmaps
- HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations
- Evaluating Attribution for Graph Neural Networks
- Self-Explainable Multi-Label Graph Neural Network for Correlated Evidence Attribution
- Node-wise Filtering in Graph Neural Networks: A Mixture of Experts Approach
- Probing GNN Explainers: A Rigorous Theoretical and Empirical Analysis of GNN Explanation Methods
- Rethinking Spectral Graph Neural Networks with Spatially Adaptive Filtering
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- Towards Training GNNs using Explanation Directed Message Passing
- Interpreting Deep Graph Convolutional Networks with Spectrum Perspective
- The Oversmoothing Fallacy: A Misguided Narrative in GNN Research
- Are we measuring oversmoothing in graph neural networks correctly?
- MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing
- A critical look at the evaluation of GNNs under heterophily: are we really making progress?