Can Fixed-Random Reservoir Weights Solve GCN Over-Smoothing Without Deep Message Passing?

Why depth breaks GCN, and what reservoir computing promised instead

GCN message passing updates each node by multiplying the normalized adjacency matrix, the feature matrix, and a trainable weight matrix, so receptive fields grow with depth [1][5]. Foundational work pushed receptive fields wider by other means: H-GCN coarsens structurally similar nodes into hyper-nodes and refines them back, reporting up to 5.9% accuracy improvement and larger gains when labels are scarce [6], while N-GCN trains multiple GCN instances over random-walk node pairs at different distances and improves Cora, Citeseer, Pubmed, and PPI baselines [7]. These are architectural detours around depth rather than fixes to depth itself.

Reservoir computing offered a different route: keep the recurrent weights fixed and random, set the spectral radius below one for contractive dynamics, and train only a readout [1][2]. GraphESN pioneered this for node embeddings, FDGNN stacked reservoir layers, Bianchi et al. added pooling, and MRGNN limited message passing to a neighborhood radius to preserve local diversity [1]. The paper's own diagnosis is that these models still lack a structured convolutional mechanism, so multi-hop aggregation is imprecise, and their contractive dynamics bias them toward short-term dependencies [1]. That is the gap RGC-Net targets.

What RGC-Net actually changes in the update rule

RGC-Net embeds a layered graph convolutional mechanism inside the reservoir framework, using fixed, non-trainable reservoir weights plus a leaky integrator that balances retained past state against new input [1]. The stated purpose of the leaky integrator is to control how much of the initial node embedding survives aggregation, preserving node individuality rather than letting embeddings collapse [1]. The paper frames this as three hypotheses: reservoir dynamics can perform graph convolution as well as or better than traditional graph convolution; the leaky integrator prevents over-smoothing; and non-trainable parameters yield faster convergence and lower resource use [1].

The trainable counterpart TRGC-Net enables a direct fixed-versus-trainable comparison, which the authors describe as the first systematic one in reservoir-based graph learning [1]. This matters because the reservoir literature has been moving the other way: AFRICO adapts input and state-feedback weights with an Extended Kalman Filter and reports up to 88% reduction in normalized mean squared error versus fixed-output-feedback echo state networks [2]. RGC-Net tests whether that adaptation is necessary for graph convolution specifically, and its answer is dataset-dependent rather than universal [1].

Classification gains, generation gains, and where the fixed reservoir wins

In graph classification the paper reports RGC-Net outperforming GCN and GAT and achieving state-of-the-art results, with faster convergence attributed to fewer parameters to optimize and a regularization effect from input-independent reservoir mapping [1]. In brain graph evolution prediction, transformer-based models (GCN-Transformer, RGC-Net-Transformer, TRGC-Net-Transformer) outperformed RBGM and EvoGraphNet across all datasets, and RBGM and EvoGraphNet hit out-of-memory errors on the larger SLIM160 dataset, which the authors attribute to memory-heavy edge-based convolution [1]. On the Simulated dataset RGC-Net achieved slightly better MAE, Frobenius distance, MAE node strength, and eigenvector centrality than TRGC-Net, suggesting the fixed reservoir captured the needed dynamics without overfitting; on EMCI-AD, TRGC-Net gained minor improvements in betweenness centrality; on SLIM160 both remained scalable, with RGC-Net achieving the lowest Frobenius distance and MAE node strength [1].

The interpretation the authors draw is a trade-off: fixed reservoirs favor efficiency, scalability, and robustness on smaller or less complex graphs, while trainable reservoirs give marginal gains on complex or irregular connectivity at higher computational cost and slower training [1]. Notably, no generative model beat the identity function on EMCI-AD, which the authors link to highly similar connectivity distributions between time points [1]. That is a useful negative result: it marks a regime where the task itself may not reward any learned mapping.

Faster convergence is not the same as cheaper computation

The paper's third hypothesis is only partially supported by its own measurements. RGC-Net-Transformer used more memory than GCN-Transformer but trained faster overall because it needed fewer epochs to converge; however, training time per epoch was slightly slower than GCN-Transformer, consistent with the higher time and memory cost of an RGC-Net layer relative to a GCN layer [1]. The authors state plainly that non-trainable reservoir weights did not demonstrate significant efficiency improvements over conventional GCN in resource consumption [1].

This distinction matters for how the result should be read. Faster wall-clock convergence can come from fewer optimization steps rather than cheaper steps, and the paper separates those two quantities rather than conflating them [1]. The authors also flag that RGC-Net is not permutation-invariant, which may affect performance on graphs with varying node orders or isomorphic structures, and that it adds two hyperparameters (leaky rate and iteration count) relative to a traditional GCN [1]. Those are architectural costs that a headline accuracy number does not capture.

Where the over-smoothing claim stops, and how it should be measured

The paper's own conclusion is scoped to the tasks it tested: graph classification and temporal brain graph generation, with future work explicitly listed for node classification and edge prediction [1]. It does not establish that fixed-random reservoir weights dominate trainable GCNs across all graph types or on very deep networks, and the authors note that further optimization is needed before deployment on resource-constrained devices [1]. The broader depth literature continues along a different line, with competing work treating deep GNN degradation through representation-evolution control and initial residual plus identity mapping rather than reservoir dynamics [3].

A separate limitation is measurement. Zhang et al. argue that leading over-smoothing measures based on Dirichlet-like energy have limitations and propose effective rank as a more robust indicator, showing theoretically that feature rank decays in message-passing GNNs [5]. They also caution that a large effective rank does not guarantee good performance, and that some smoothing can be beneficial [5]. RGC-Net reports mitigated over-smoothing without the supplied material specifying which metric was used, so the claim should be read as consistent with the paper's own evaluation rather than as an independently validated reduction in a canonical over-smoothing measure [1][5]. Validation on other graph domains is also thin in the supplied evidence: R-GNN shows recurrent graph architectures working on oracle bone inscription font classification with 88.2% Top-1 accuracy and residual GRUs fusing local and global features [4], but that is a different task family and does not test RGC-Net's specific fixed-reservoir design.

About These Sources

This research page is built on 7 studies (6 peer-reviewed, 1 preprint) — published from 2018 to 2026, 5 from 2024 or later, collectively cited 514 times — selected as the most relevant from 13 studies that passed quality screening, drawn from 111 papers retrieved from a database of over 500 million.

Sources used in this answer