Research map / Graph learning / Knowledge and temporal graphs
Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph
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 Delay-DSGN, a spiking dynamic graph neural network that adds a learnable Gaussian synaptic delay kernel to neighborhood aggregation to model delayed propagation and reduce historical forgetting, evaluated on node classification over DBLP, Tmall and Patent.
Problem
Existing SNN-based dynamic graph methods fail to model latency in information propagation and historical information forgetting when learning node representations on evolving graphs.
Approach
Introduce a learnable synaptic delay via a Gaussian delay convolution kernel (per-synapse delay and weight) into SNN temporal modeling, so historical spikes influence future node representations; add a theoretical condition relating kernel standard deviation and size to gradient stability.
Key results
- On DBLP at 80% training ratio, Delay-DSGN reaches 76.87% Mi-F1, 2.21 points above second-best Dy-SIGN.
- On Tmall at 80% training ratio, Mi-F1 of 67.88%, 4.56 points above TGAT.
- On Patent, best Ma-F1 of 84.20% at 80% training ratio, though gains over SpikeNet and Dy-SIGN are small (about 0.3 points).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a learnable delay mechanism | Delay-DSGN · DBLP, 80% training | Mi-F1: 74.96±0.2 -> 76.87±0.5 | Helped |
| Add a learnable delay mechanism | Delay-DSGN · Tmall, 80% training | Mi-F1: 66.10 -> 67.88±0.4 | Helped |
| Add a learnable delay mechanism | Delay-DSGN · Patent, 80% training | Ma-F1: 83.91±0.2 -> 84.20±0.1 | Helped |
| Use fixed random delays instead of learnable delays (ablation) | Delay-DSGN · DBLP, Tmall, and Patent | Ma-F1 and Mi-F1 | Hurt |
| Remove the delay module (ablation) | Delay-DSGN · DBLP, Tmall, and Patent | Ma-F1 and Mi-F1 | Hurt |
| Increase the maximum delay time dm (ablation) | Delay-DSGN · DBLP and Tmall | Ma-F1 and Mi-F1 | Helped |
| Use a moderate Gaussian kernel standard deviation σ (ablation) | Delay-DSGN · DBLP and Tmall | Ma-F1 and Mi-F1 | Helped |
| Use too small or too large Gaussian kernel standard deviation σ (ablation) | Delay-DSGN · DBLP and Tmall | Ma-F1 and Mi-F1 | Hurt |
Surprising findings
- On Patent, Delay-DSGN's gains over Dy-SIGN and SpikeNet are marginal (e.g., 84.20 vs 83.91 Ma-F1), within reported standard deviations in some cases.
Limitations
- Gains over strong SNN baselines on Patent are small.
- Baselines HTNE and M2DNE fail to run on Patent due to cost, limiting comparisons at scale.
- Theoretical gradient bound relies on assumptions about Gaussian kernel behavior, and the full proof is deferred to an appendix.
Experimental setup
Datasets: DBLP Tmall Patent
Benchmarks: DBLP Tmall Patent
Metrics: Macro-F1 Micro-F1
Baselines: DeepWalk Node2Vec HTNE M2DNE EvolveGCN TGAT SpikeNet Dy-SIGN
Builds on: SpikeNet Dy-SIGN LIF neuron
Closest papers by meaning
- Dynamic Spiking Graph Neural Networks (AAAI 2024 · similarity 0.92)
- Dynamic Reactive Spiking Graph Neural Network (AAAI 2024 · similarity 0.90)
- Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning (ICML 2024 · similarity 0.90)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.89)
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.88)
- PRES: Toward Scalable Memory-Based Dynamic Graph Neural Networks (ICLR 2024 · similarity 0.87)
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.