SlepNet: Spectral Subgraph Representation Learning for Neural Dynamics

Siddharth Viswanath, Rahul Singh, Yanlei Zhang, J. Adam Noah, Joy Hirsch, Smita KrishnaswamyarXiv 2025 · DOI · arXiv

Introduction

Most Graph Neural Networks (GNNs) rely on the Graph Fourier Transform (GFT) to perform spectral filtering. However, GFT harmonics are global—they spread across the entire graph. This makes them "blurry" when trying to represent signals that are localized to specific regions, such as activity in the visual cortex during a movie or a traffic jam on a specific highway segment. While Wavelet transforms offer some localization, they often suffer from "information leakage" at subgraph boundaries. SlepNet addresses this by providing a principled way to learn representations that are simultaneously band-limited (smooth in frequency) and space-limited (confined to a specific subgraph) without predefined boundaries.

Methodology

SlepNet introduces a two-part architecture to solve the localization problem:

  1. Attention-based Mask Learning: SlepNet uses an attention mechanism to learn a mask ($m$) that identifies meaningful, contiguous regions. This is achieved by clustering nodes using spectral clustering and applying the mask to these clusters.
  2. Slepian Spectral Filtering: The model computes Graph Slepians—harmonics mathematically optimized to maximize energy concentration within the masked subgraph while remaining spectrally band-limited.
  3. Neural Eigenmapping: To scale to large graphs, the authors propose a "Neural Eigenmapper"—a supervised regression network that learns to approximate Slepian harmonics, bypassing the expensive $O(N^3)$ eigendecomposition.
Overview of the SlepNet architecture, illustrating the transition from global graph data to localized Slepian spectral filtering.

Experimental Evaluation

The model was benchmarked across diverse spatiotemporal domains including Neuroscience (fMRI datasets like OCD-PVDM, OCD-RA, and ASD-ABIDE) and Traffic Dynamics (PEMS03 and PEMS07).

Key Results

  • Superior Accuracy: On the OCD-PVDM dataset, SlepNet-I achieved 84.70% accuracy, significantly outperforming standard Spectral GCN (62.62%) and GraphSAGE (55.01%).
  • Traffic Prediction: In the PEMS07 traffic dataset, SlepNet reached 70.16% accuracy for day-of-week prediction, while standard GCNs hovered between 18-30%.
  • Scalability: The Neural Eigenmapping approach maintained stable runtimes as graph sizes increased, overcoming the computational bottlenecks of classical methods.

Discussion and Key Insights

By focusing on subgraphs, SlepNet ignores noise from inactive parts of the graph. In fMRI, this allows the model to focus on specific brain circuits relevant to psychiatric conditions. Furthermore, SlepNet represents signal evolution as a trajectory in latent space. These trajectories exhibit higher curvature and resolution than those from standard GNNs, capturing the nuances of system state transitions.

Visualization of Slepian harmonics concentrated on specific subgraphs, demonstrating localized energy distribution.

The learned masks also provide interpretability, successfully identifying biological regions associated with OCD, such as the parietal lobe and dorsolateral prefrontal cortex.

Limitations and Future Work

Current limitations include a fixed spectral bandwidth and a focus on discriminative tasks. Future work may involve making the frequency band learnable, extending the framework to generative modeling of neural dynamics, and integrating recurrent or transformer-based layers for enhanced temporal modeling.

Summary updated

Abstract

Graph neural networks have been useful in machine learning on graph-structured data, particularly for node classification and some types of graph classification tasks. However, they have had limited use in representing patterning of signals over graphs. Patterning of signals over graphs and in subgraphs carries important information in many domains including neuroscience. Neural signals are spatiotemporally patterned, high dimensional and difficult to decode. Graph signal processing and associated GCN models utilize the graph Fourier transform and are unable to efficiently represent spatially or spectrally localized signal patterning on graphs. Wavelet transforms have shown promise here, but offer non-canonical representations and cannot be tightly confined to subgraphs. Here we propose SlepNet, a novel GCN architecture that uses Slepian bases rather than graph Fourier harmonics. In SlepNet, the Slepian harmonics optimally concentrate signal energy on specifically relevant subgraphs that are automatically learned with a mask. Thus, they can produce canonical and highly resolved representations of neural activity, focusing energy of harmonics on areas of the brain which are activated. We evaluated SlepNet across three fMRI datasets, spanning cognitive and visual tasks, and two traffic dynamics datasets, comparing its performance against conventional GNNs and graph signal processing constructs. SlepNet outperforms the baselines in all datasets. Moreover, the extracted representations of signal patterns from SlepNet offers more resolution in distinguishing between similar patterns, and thus represent brain signaling transients as informative trajectories. Here we have shown that these extracted trajectory representations can be used for other downstream untrained tasks. Thus we establish that SlepNet is useful both for prediction and representation learning in spatiotemporal data.
BibTeX
@misc{viswanath2025slepnet,
  author = {Siddharth Viswanath and Rahul Singh and Yanlei Zhang and J. Adam Noah and Joy Hirsch and Smita Krishnaswamy},
  title = {{SlepNet: Spectral Subgraph Representation Learning for Neural Dynamics}},
  year = {2025},
  eprint = {2506.16602},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.16602},
}
Figure