SiGNN: A Spike-induced Graph Neural Network for Dynamic Graph Representation Learning

TL;DR

This work addresses dynamic graph representation learning by leveraging the temporal processing capabilities of Spiking Neural Networks (SNNs) and fusing them with Graph Neural Networks (GNNs) through a Temporal Activation (TA) mechanism. The proposed SiGNN framework also introduces Multiple Time Granularities (MTG) to capture multiscale temporal dynamics, enabling rich spatiotemporal node representations. Key contributions include the TA mechanism that integrates SNN dynamics without binarizing features, the Bidirectional LIF (BLIF) neuron for richer temporal signaling, and MTG-aware sampling and aggregation for multiscale temporal modeling. Experiments on three real-world datasets show that SiGNN consistently outperforms state-of-the-art baselines in temporal node classification, with insights into how spike dynamics relate to graph evolution and how MTG and TA contribute to performance.

Abstract

In the domain of dynamic graph representation learning (DGRL), the efficient and comprehensive capture of temporal evolution within real-world networks is crucial. Spiking Neural Networks (SNNs), known as their temporal dynamics and low-power characteristic, offer an efficient solution for temporal processing in DGRL task. However, owing to the spike-based information encoding mechanism of SNNs, existing DGRL methods employed SNNs face limitations in their representational capacity. Given this issue, we propose a novel framework named Spike-induced Graph Neural Network (SiGNN) for learning enhanced spatialtemporal representations on dynamic graphs. In detail, a harmonious integration of SNNs and GNNs is achieved through an innovative Temporal Activation (TA) mechanism. Benefiting from the TA mechanism, SiGNN not only effectively exploits the temporal dynamics of SNNs but also adeptly circumvents the representational constraints imposed by the binary nature of spikes. Furthermore, leveraging the inherent adaptability of SNNs, we explore an in-depth analysis of the evolutionary patterns within dynamic graphs across multiple time granularities. This approach facilitates the acquisition of a multiscale temporal node representation.Extensive experiments on various real-world dynamic graph datasets demonstrate the superior performance of SiGNN in the node classification task.

Figures (7)

• Figure 1: SNNs are widely employed for efficiently processing time-series data. Given that dynamic graph data also exhibit temporal dependencies, SNNs hold significant potential for capturing the temporal dynamics in evolving graphs.
• Figure 2: The framework of SiGNN. SiGNN consists of three fundamental steps: MTG-aware graph sampling , Spike-induced graph aggregation and temporal embedding aggregation.
• Figure 3: The schematic diagram of the BLIF neuron model. Figure (a) illustrates the internal computational flow of BLIF neurons at time $t$, while Figure (b) represents the variation of membrane voltage over time and the bidirectional spike firing characteristics in BLIF neurons.
• Figure 4: The Macro-F1 and Micro-F1 metrics for three methods: TA-BLIF, TA-LIF, and without TA mechanism. The x-axis represents various training ratios.
• Figure 5: The performance of models with varying numbers of time granularity channels. The x-axis represents the count of introduced time granularities.