Research map / Graph learning / GNN architectures
Dynamic Spiking Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dy-SIGN, a spiking dynamic graph neural network that adds an information-compensation channel from early to last layer and uses implicit differentiation at the equilibrium state for memory-efficient training on dynamic node classification.
Problem
Applying spiking neural networks to dynamic graph node classification suffers from information loss due to binary spikes and high memory cost of backpropagation through time across graph snapshots and SNN latency steps.
Approach
Bridges shallow-layer features directly to the final layer (information compensation) and replaces BPTT with implicit differentiation at a fixed point, extended to dynamic graphs by propagating the equilibrium state across time steps instead of hidden states.
Key results
- Dy-SIGN achieves Macro-F1 of 74.67% on DBLP at 80% training ratio, comparable to SpikeNet (74.64%).
- Dy-SIGN uses about half the memory of SpikeNet on DBLP according to Figure 2(a).
- Dy-SIGN outperforms SpikeNet in half of the Macro-F1 and Micro-F1 settings, and is slightly faster than SpikeNet per epoch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SpikeNet with Dy-SIGN | Dynamic node classification · DBLP, Tmall, and Patent | Macro-F1: DBLP: 70.88±0.4 -> 70.94±0.1; 71.98±0.3 -> 72.07±0.1; 74.64±0.5 -> 74.67±0.5. Tmall: 58.84±0.4 -> 57.48±0.1; 61.13±0.8 - | Mixed |
| Replace SpikeNet with Dy-SIGN | Dynamic node classification · DBLP, Tmall, and Patent | Micro-F1: DBLP: 71.98±0.3 -> 71.90±0.1; 72.35±0.8 -> 72.61±0.4; 74.86±0.5 -> 74.96±0.2. Tmall: 63.52±0.7 -> 62.93±0.3; 64.84±0.4 - | Mixed |
| Use Dy-SIGN's variation training method instead of BPTT-based training | Dy-SIGN · DBLP | memory consumption: about only half | Helped |
| Use Dy-SIGN's variation training method instead of BPTT-based training | Dy-SIGN · Training | training time | Helped |
| Vary λ from smaller to larger values (ablation) | Dy-SIGN · Dynamic node classification | accuracy | Mixed |
Surprising findings
- Dy-SIGN does not uniformly beat SpikeNet; it is slightly lower on Tmall and Patent in several settings.
Limitations
- Gains over SpikeNet are marginal and inconsistent across settings.
- Gradient computation simplified by zeroing the Heaviside derivative, an approximation not fully justified empirically.
- Only node classification on three datasets evaluated.
Experimental setup
Datasets: DBLP Tmall Patent
Benchmarks: DBLP Tmall Patent
Metrics: Macro-F1 Micro-F1 memory consumption training time
Baselines: DeepWalk Node2Vec HTNE M2DNE DyTriad MPNN JODIE EvolveGCN TGAT SpikeNet
Builds on: SpikeNet Deep Equilibrium Models MDEQ EvolveGCN BPTT with surrogate gradients
Closest papers by meaning
- Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph (ICML 2025 · similarity 0.92)
- Dynamic Reactive Spiking Graph Neural Network (AAAI 2024 · similarity 0.91)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.88)
- Sign Gradient Descent-based Neuronal Dynamics: ANN-to-SNN Conversion Beyond ReLU Network (ICML 2024 · similarity 0.88)
- Node-Time Conditional Prompt Learning in Dynamic Graphs (ICLR 2025 · similarity 0.87)
- Signed Graph Neural Ordinary Differential Equation for Modeling Continuous-Time Dynamics (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.