Title: DynSpike: a spiking neural network for effective long-term temporal dependency capture in dynamic graphs

Authors: Zijing Yuan; Tianfang Lu; Masaaki Omura; Shangce Gao

Addresses: Faculty of Engineering, University of Toyam, Toyama-shi, Japan ' Faculty of Engineering, University of Toyam, Toyama-shi, Japan ' Faculty of Engineering, University of Toyam, Toyama-shi, Japan ' Faculty of Engineering, University of Toyam, Toyama-shi, Japan

Abstract: Dynamic graph learning is crucial for modelling systems with evolving node features and graph structures. Spiking neural networks offer low-power solutions but struggle with long-term dependencies and multi-scale dynamics. To address these, we propose DynSpike, a novel spiking neural network framework for dynamic graph processing. DynSpike introduces a spatio-temporal multi-scale fusion strategy, modelling local/global structures and short/long-term dependencies through parallel spatio-temporal branches. These branches are processed by the aggregation-spiking-attention module, combining neighbourhood aggregation, spike-driven encoding, and multi-head attention for fine-grained node representations. Experiments on benchmark datasets show DynSpike outperforms existing methods in node classification, effectively capturing multi-scale spatio-temporal patterns with high efficiency.

Keywords: deep learning; graph neural network; spiking neural networks; dynamic graph data.

DOI: 10.1504/IJBIC.2026.155425

International Journal of Bio-Inspired Computation, 2026 Vol.28 No.1, pp.12 - 23

Received: 30 Sep 2025
Accepted: 25 Nov 2025
Published online: 31 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article