Spectrum-aware Spiking Dynamics for Graph Contrastive Learning

Beibei Wang, Zhiyan Zhang, Bo Jiang, Yapeng Li, Qin Xu, wanli lyu

Keywords: Graph Contrastive Learning, Graph Spectral Information, Spiking Neural Networks, Membrane Learning

Abstract: Graph Contrastive Learning (GCL) typically relies on Graph Neural Networks (GNNs) for full-precision representation learning, which results in high computational overhead and energy consumption. Recently, integrating Spiking Neural Networks (SNNs) with GNNs has emerged as a promising energy-efficient alternative. However, existing approaches often treat spiking neurons merely as binary encoders to produce 1-bit representations, ignoring the rich structural information inherent in graph data. Also, the usage of a fixed initial membrane potential (IMP), which is usually set to 0, restricts the diversity of spiking patterns and limits the expressive power of spiking neurons. To address these issues, we propose a novel Spectrum-enhanced Spiking Graph Contrastive Learning (S$^2$GCL) framework by integrating graph spectral information into spiking dynamics. Specifically, we first develop a novel Spectrum-aware Membrane Potential (SaMP) mechanism for SNNs by injecting eigenvalue-based biases into membrane potential learning to capture global graph structure and enhance SNN's expressive power. Then, we introduce an Overlapped Channel Grouping (OCG) strategy to construct sequence spikes for the graph and simultaneously reinforce correlations in spike trains based on overlapped feature channels. Finally, we adopt the dual-level contrastive objective to achieve both node-wise and channel-wise alignments. Extensive experiments on several benchmark datasets show the effectiveness of our proposed S$^2$GCL. The code of our method will be released upon acceptance.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 22807