MRGNN: Multiresolution Reservoir Graph Neural Network

• MRGNN is a graph-level architecture that combines contractive reservoir dynamics with topological pooling for consistent fixed-point embeddings.
• It alternates reservoir updates and graph coarsening to balance computational efficiency with near-baseline accuracy on diverse graph tasks.
• Spectral reinterpretations, including a Fairing-based variant, address over-smoothing by enabling selective filtering and random-walk redundancy control.

Multiresolution Reservoir Graph Neural Network (MRGNN), also described as a pyramidal Reservoir Graph Neural Network, is a graph-level learning architecture that interleaves reservoir-computing layers with topological graph-pooling layers so that node representations are repeatedly updated, coarsened, and finally aggregated into a graph embedding. In the 2021 formulation, each reservoir layer is defined as a contractive dynamical system whose node states converge to a unique fixed point, while each pooling layer reduces the support graph and the feature support, yielding a multiresolution or pyramidal computation that is explicitly analyzed in terms of convergence and complexity (Bianchi et al., 2021). A later spectral reinterpretation revisits the reservoir definition through Laplacian filtering and introduces a Fairing-based variant designed to control smoothing more selectively, with an associated random-walk interpretation of redundancy (Bison et al., 17 Jul 2025).

1. Architectural organization

The core MRGNN architecture is organized into $L$ blocks. For each block $l=1,\dots,L-1$, the model alternates two operations: an RC layer that embeds every node via a contractive dynamical system, and a topological pooling layer that produces a coarsened graph with fewer nodes, fewer edges, and a smaller spectral radius. The final $L$-th block contains only an RC layer. The remaining node states are then aggregated by summation into a single $H$-dimensional graph embedding,

$\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$

This graph embedding is the representation used for downstream graph-level prediction (Bianchi et al., 2021).

At layer $l$, the hidden-state matrix is denoted $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$, the input features are $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$, and the normalized adjacency is $\tilde A^{(l)}$. The iterated update is

$$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$

For $l=1,\dots,L-1$0, $l=1,\dots,L-1$1 is the original feature dimension; for $l=1,\dots,L-1$2, $l=1,\dots,L-1$3. The fixed-point state $l=1,\dots,L-1$4 becomes the input $l=1,\dots,L-1$5 to the next block. This arrangement gives the model its multiresolution character: node dynamics are computed at one scale, then the graph is coarsened, and the process repeats on a smaller graph.

A later exposition of MRGNN reframes the reservoir as a fixed, non-learned graph filter of depth $l=1,\dots,L-1$6, using the symmetric normalized Laplacian

$l=1,\dots,L-1$7

with $l=1,\dots,L-1$8. In that view, the reservoir step is written

$l=1,\dots,L-1$9

and the intermediate embeddings are concatenated into

$L$0

The same source describes a simple two-layer MLP readout for graph-level classification or regression, which indicates that later work uses the MRGNN label both for the original multiresolution fixed-point construction and for a spectral reservoir interpretation built around the same general reservoir-on-graphs paradigm (Bison et al., 17 Jul 2025).

2. Reservoir dynamics and fixed-point theory

In the fixed-point formulation, a reservoir layer updates node states by

$L$1

where $L$2 is the symmetrically normalized adjacency matrix of the current graph, $L$3 is the node-state matrix at iteration $L$4, $L$5 is the input feature matrix, $L$6, $L$7, and $L$8 is broadcast to all nodes. The summary notes that, in practice, $L$9 is often omitted or absorbed; it also states that the paper uses $H$0 for $H$1 of size $H$2 and a hidden-scaling matrix for $H$3 of size $H$4 (Bianchi et al., 2021).

The central theoretical property is contraction. If

$H$5

then the map is $H$6-Lipschitz with

$H$7

where $H$8 is the spectral norm of $H$9 and $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$0 is the largest-magnitude eigenvalue of $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$1. When $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$2, described as the “Graph Embedding Stability” condition, $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$3 is a contraction, so a unique fixed point $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$4 exists and the iteration converges geometrically from any initialization. In the implementation pipeline, the initialization is $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$5 (Bianchi et al., 2021).

The paper also provides an explicit bound on the convergence time. If the layer is iterated until $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$6, then the number of iterations $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$7 satisfies

$\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$8

where $\phi(G)=\sum_{n=1}^{N_L} x_n^{(L)*}.$9 is the norm after one step. The stated consequence is that $l$0 grows as $l$1, that is, as $l$2 or $l$3 grow. This establishes a direct connection between graph topology, reservoir scaling, and the runtime needed to obtain the fixed-point representation (Bianchi et al., 2021).

The original graph-level readout after the last reservoir layer is the row-wise summation of $l$4, followed by a linear ridge-regression classifier trained on $l$5. A common misconception is that the reservoir is merely an initialization stage for a conventional trainable GNN. In the reported formulation, the reservoir itself is the principal feature extractor, while the trainable component is the final readout. The later spectral reinterpretation preserves this reservoir-centered perspective, although it reformulates the propagation operator and the readout construction (Bison et al., 17 Jul 2025).

3. Graph coarsening and the multiresolution trade-off

The multiresolution component is defined through an assignment or selection matrix $l$6 with $l$7. The coarsened adjacency and feature matrices are

$l$8

Applying the coarsening operator successively reduces $l$9, $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$0, and typically the spectral radius $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$1. The paper’s theoretical argument is that this affects not only storage and per-iteration cost, but also the contraction speed of the reservoir updates (Bianchi et al., 2021).

Three pooling instantiations are listed.

Pooling method	Mechanism	Reported note
Graclus	Unsupervised spectral clustering into pairs, then average-pool	May introduce “fake” nodes
NMF pooling	$X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$2 with non-negative $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$3, then aggregate by $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$4	Can yield dense coarsened graphs
Node-Decimation (NDP)	Drop $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$5half the nodes via max-cut partition, reconnect via Kron-reduction, and sparsify	Reduces $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$6, $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$7, and $X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$8 most effectively

The computational analysis quantifies the benefit. Each iteration of a reservoir layer costs

$X^{(l)}[t]\in\mathbb{R}^{N_l\times H}$9

corresponding to sparse $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$0-times-$U^{(l)}\in\mathbb{R}^{N_l\times F_l}$1 and dense $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$2. Summed over $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$3 iterations, one layer costs

$U^{(l)}\in\mathbb{R}^{N_l\times F_l}$4

Because pooling reduces $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$5, $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$6, and typically $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$7, it reduces the number of matrix multiplications and accelerates contraction. The theoretical point is therefore stronger than ordinary downsampling: pooling is analyzed as a mechanism that alters the convergence properties of the reservoir dynamics themselves (Bianchi et al., 2021).

This also clarifies an important trade-off. Pooling is not presented as uniformly accuracy-improving. Rather, it provides what the paper describes as a principled trade-off between accuracy and complexity. A plausible implication is that MRGNN is best understood as a family of operating points on an accuracy–efficiency frontier, rather than as a single architecture with a fixed optimum.

4. Training protocol, datasets, and empirical behavior

The reported training and inference workflow is explicit. First, pooling assignments $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$8 and coarsened $U^{(l)}\in\mathbb{R}^{N_l\times F_l}$9 are precomputed for each block using Graclus, NMF, or NDP. For each input graph $\tilde A^{(l)}$0, the reservoir iteration is run from $\tilde A^{(l)}$1 until $\tilde A^{(l)}$2 or $\tilde A^{(l)}$3; the result is then pooled by $\tilde A^{(l)}$4 to obtain the next layer’s input features, and the process is repeated. After the last RC layer, the node states are summed to form $\tilde A^{(l)}$5. A ridge-regression readout is fitted on $\tilde A^{(l)}$6. At inference time, the same embedding pipeline is used and only readout training is omitted (Bianchi et al., 2021).

The evaluation covers two synthetic datasets and a set of established benchmarks. The synthetic datasets are B-easy and B-hard, described as two new synthetic “structure-+-feature informative only jointly” datasets with 3 classes each and $\tilde A^{(l)}$7 graphs. The benchmark suite comprises ten TUD datasets spanning molecules, including Mutagenicity and Enzymes, and social networks, including IMDB-BIN and Reddit. The baseline is “No-pool,” defined as the same depth $\tilde A^{(l)}$8 but without pooling (Bianchi et al., 2021).

The key empirical findings are differentiated rather than uniform. No-pool achieves slightly higher accuracy, especially on small graphs, but pyramidal models run up to $\tilde A^{(l)}$9 faster on large graphs. Among the pooling methods, NDP gives the best accuracy/time trade-off, especially on Reddit-2K/5K, because it reduces $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$0, $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$1, and $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$2 most effectively. The accuracy/time ratio, defined as accuracy divided by train-plus-test seconds, is reported as consistently higher for pyramidal models, and NDP-based MRGNN is best overall. Qualitative LDA visualizations of $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$3 show only mild degradation in class separability under pooling (Bianchi et al., 2021).

These results address a frequent misconception that the pyramidal design is valuable only when it preserves accuracy exactly. The reported evidence instead supports a more specific claim: on large graphs, the principal benefit is improved efficiency with near-baseline accuracy, whereas on small graphs the unpooled baseline can still be slightly more accurate. The architecture is thus empirically positioned as an efficiency-oriented reservoir GNN rather than as a uniformly dominant alternative.

5. Spectral reinterpretation and the Fairing-based reservoir

A 2025 paper revisits the reservoir in MRGNN as a spectral reservoir model and proposes a variant based on a Fairing algorithm originally introduced in surface design in computer graphics (Bison et al., 17 Jul 2025). The motivation is over-smoothing: repeated application of low-pass graph operators drives node features toward low-frequency Laplacian components, making node embeddings increasingly indistinguishable.

In the spectral formulation, with $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$4, any polynomial filter $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$5 can be written as

$$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$6

Taubin’s Fairing algorithm alternates a shrinking low-pass step with an unshrinking high-pass component. For spectral frequency $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$7, one full Fairing cycle with $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$8 shrink steps of parameter $$X^{(l)}[t+1] = \tanh\!\bigl( \tilde A^{(l)} X^{(l)}[t] W^{(l)} + U^{(l)} W_{\mathrm{in}}^{(l)} \bigr).$$9 and $l=1,\dots,L-1$00 unshrink steps of parameter $l=1,\dots,L-1$01 gives

$l=1,\dots,L-1$02

The source characterizes this as a pass-band filter that smooths intermediate frequencies while preserving the global norm and avoiding the monotonic decay of all but the zero-frequency component.

For a single Fairing cycle with parameter $l=1,\dots,L-1$03, the operator is

$l=1,\dots,L-1$04

and iterating it $l=1,\dots,L-1$05 times yields

$l=1,\dots,L-1$06

More generally, alternating shrink and unshrink gives

$l=1,\dots,L-1$07

This makes the spectral selectivity of the reservoir explicit (Bison et al., 17 Jul 2025).

The same paper derives a random-walk interpretation. With transition matrix $l=1,\dots,L-1$08,

$l=1,\dots,L-1$09

so

$l=1,\dots,L-1$10

Applying $l=1,\dots,L-1$11 therefore mixes $l=1,\dots,L-1$12-step walk contributions with signed binomial weights, and the source states that these signed sums heavily over-count tottering walks. Fairing instead uses

$l=1,\dots,L-1$13

Under this filter, the probability that a random vacuous/non-vacuous walk has exactly $l=1,\dots,L-1$14 non-vacuous edges is

$l=1,\dots,L-1$15

The stated interpretation is that $l=1,\dots,L-1$16 near $l=1,\dots,L-1$17 suppresses moves, $l=1,\dots,L-1$18 near $l=1,\dots,L-1$19 emphasizes $l=1,\dots,L-1$20-hop walks, and intermediate $l=1,\dots,L-1$21 spreads weight across walk lengths, thereby controlling both over-smoothing and expressive reach (Bison et al., 17 Jul 2025).

6. Configuration guidelines, comparative results, and open problems

The technical summary accompanying the original model provides concrete guidelines for adaptation to new graph tasks. For reservoir hyperparameters, hidden size $l=1,\dots,L-1$22–$l=1,\dots,L-1$23 is reported to work well, with the caveat that larger $l=1,\dots,L-1$24 increases expressivity but also the per-update cost $l=1,\dots,L-1$25. The spectral radius $l=1,\dots,L-1$26 is recommended in $l=1,\dots,L-1$27 to satisfy contraction $l=1,\dots,L-1$28, and input or hidden scalings $l=1,\dots,L-1$29 are suggested in $l=1,\dots,L-1$30. The convergence threshold $l=1,\dots,L-1$31 and $l=1,\dots,L-1$32 are described as robust defaults. For pooling, NDP is recommended for large, sparse graphs; the sparsification parameter is tuned in $l=1,\dots,L-1$33; Graclus is to be avoided on very large graphs because it may introduce fake nodes; NMF is to be avoided if $l=1,\dots,L-1$34 because it can yield dense coarsened graphs; and halving $l=1,\dots,L-1$35 per layer is recommended to balance abstraction against information loss. For readout, a linear or ridge-regression model is considered sufficient when $l=1,\dots,L-1$36 is high-quality, while a small MLP is suggested for highly nonlinear tasks (Bianchi et al., 2021).

The 2025 Fairing study reports exploratory graph-classification results on four benchmarks, using the original MRGNN defaults and no further model selection. The setup uses reservoir depth $l=1,\dots,L-1$37, hidden readout dimensions $l=1,\dots,L-1$38, and learning rate in $l=1,\dots,L-1$39, with dropout, weight-decay, and batch-size as in the original. Results are reported with 10-fold cross-validation $l=1,\dots,L-1$40 5 runs, as mean $l=1,\dots,L-1$41 std (%) (Bison et al., 17 Jul 2025).

Dataset	MRGNN ($l=1,\dots,L-1$42)	Fairing reservoir
PTC	$l=1,\dots,L-1$43	$l=1,\dots,L-1$44
NCI1	$l=1,\dots,L-1$45	$l=1,\dots,L-1$46
PROTEINS	$l=1,\dots,L-1$47	$l=1,\dots,L-1$48
ENZYMES	$l=1,\dots,L-1$49	$l=1,\dots,L-1$50

The paper states that, despite no hyperparameter-search for Fairing, accuracies are comparable to MRGNN, confirming that controlled spectral smoothing can match reservoir performance. It further identifies several open directions: learnable Fairing parameters $l=1,\dots,L-1$51 per layer or per graph; adaptive choice of $l=1,\dots,L-1$52 based on graph spectra; extension to directed or weighted graphs through asymmetric Laplacians; combination with residual or jumping-knowledge readouts; and theoretical bounds on mixing times and representational capacity as functions of $l=1,\dots,L-1$53 (Bison et al., 17 Jul 2025).

Taken together, these sources define MRGNN as a reservoir-centric, multiresolution graph architecture in which coarsening is integral to both efficiency and dynamical stability. The original formulation emphasizes fixed-point reservoir embeddings under contraction and a mathematically grounded accuracy–complexity trade-off, while the later spectral work suggests that the same framework can be interpreted through graph signal processing, pass-band filtering, and random-walk redundancy control.