Spiking Heterogeneous Graph Attention Networks

Conference: AAAI 2026 arXiv: 2601.02401 Code: https://github.com/QianPeng369/SpikingHAN Area: Graph Learning / Spiking Neural Networks Keywords: Heterogeneous Graphs, Spiking Neural Networks, Meta-path, Graph Attention, Energy-Efficient Computing

TL;DR

This paper proposes SpikingHAN, the first framework to introduce Spiking Neural Networks (SNNs) into heterogeneous graph learning. It employs a single-layer graph convolution with shared parameters to aggregate meta-path-based neighborhood information, fuses multiple meta-path semantics via semantic-level attention, and encodes the resulting representations into 1-bit binary spike sequences. SpikingHAN achieves competitive node classification performance on three datasets with fewer parameters, faster inference, and lower energy consumption.

Background & Motivation

1. Background: Heterogeneous Graph Neural Networks (HGNNs) such as HAN effectively handle multi-type nodes and edges, capturing rich structural and semantic information. SNNs have demonstrated success in computer vision and homogeneous graph learning (e.g., SpikingGCN, SpikeGCL).
2. Limitations of Prior Work: HGNNs tend to be architecturally complex — HAN assigns an independent attention module to each meta-path, causing parameters, memory, and computational cost to scale significantly with the number of meta-paths, making deployment on resource-constrained devices difficult.
3. Key Challenge: Heterogeneous graph learning requires effective modeling of heterogeneous information, yet existing methods incur prohibitive computational overhead. The binary, sparse communication of SNNs is naturally suited for low-energy scenarios but has not been explored on heterogeneous graphs.
4. Goal: Design an energy-efficient heterogeneous graph neural network capable of operating under resource-constrained conditions.
5. Key Insight: Replace HAN's multiple independent aggregation modules with a single-layer GCN with shared parameters to simplify the model, then encode heterogeneous information into spike sequences via SNNs to achieve binarization.
6. Core Idea: Simplified heterogeneous graph convolution with shared parameters + SNN spike encoding = energy-efficient heterogeneous graph learning.

Method

Overall Architecture

The framework consists of three components: (1) Shared Convolution and Attention — a single-layer GCN with shared parameters aggregates neighborhood information for each meta-path, and semantic-level attention fuses multi-meta-path semantics; (2) Fully Connected Layer and SNN — a linear transformation produces the input current for the SNN, which performs integrate-and-fire-reset operations to generate spike sequences; (3) Spike Aggregation and Prediction — average pooling over multi-timestep spike outputs yields firing rates as classification predictions.

Key Designs

1. Single-Layer Graph Convolution with Shared Parameters

   • Function: Aggregate neighborhood information for each meta-path with minimal model complexity via shared convolution.
   • Mechanism: Unlike HAN, which assigns an independent aggregation module to each meta-path \(\Phi_p\), SpikingHAN applies a single shared weight matrix \(W_1\) across all meta-paths using standard GCN convolution: \(h_i^{\Phi_p} = \sigma(\sum_{j \in N_i^{\Phi_p}} h_j^0 \cdot W_1 / \sqrt{\tilde{D}_i \cdot \tilde{D}_j})\). This produces \(P\) sets of meta-path-specific node embeddings \(\{h^{\Phi_1}, ..., h^{\Phi_P}\}\).
   • Design Motivation: Independent modules cause parameter count to grow linearly with the number of meta-paths and are prone to overfitting. Shared parameters substantially reduce complexity while preserving expressiveness — semantic differences across meta-paths are distinguished by the subsequent attention mechanism.

2. Semantic-Level Attention and SNN Encoding

   • Function: Adaptively learn the importance of different meta-paths, fuse them, and encode the result into spike representations.
   • Mechanism: Semantic attention weights are computed as \(\beta_{\Phi_p} = \text{softmax}(\frac{1}{|V|}\sum_i q^T \cdot \tanh(h_i^{\Phi_p} \cdot W_2 + b))\), and the weighted aggregation is \(H = \sum_p \beta_{\Phi_p} \cdot h^{\Phi_p}\). This is then passed into a simplified fully connected layer without nonlinear activation or bias, \(H \cdot W_3\), serving as the SNN input current. The SNN uses the PLIF (Parametric LIF) model with a learnable membrane time constant \(\tau_m\), generating spikes through integrate-and-fire-reset dynamics. The final classification uses the average spike output over \(T\) timesteps: \(\hat{y} = \frac{1}{T}\sum_t \Theta(V^t)\).
   • Design Motivation: The binary communication of SNNs (0/1 spikes) substantially reduces computational and memory overhead compared to floating-point representations. The simplified fully connected layer (no nonlinearity, no bias) is motivated by findings from LightGCN — depth is not a critical factor in graph-based learning.

Loss & Training

Semi-supervised learning with cross-entropy loss; backpropagation through time via surrogate gradients.

Key Experimental Results

Main Results

Node classification on three heterogeneous graph datasets (Mi-F1, 20% training ratio):

Method	Type	DBLP	ACM	IMDB
GAT	Homogeneous GNN	91.4	91.1	58.5
SpikingGCN	Homogeneous SNN	88.7	91.2	51.3
HAN	Heterogeneous GNN	93.1	92.8	61.3
PHGT	Heterogeneous GNN	93.7	93.4	63.2
SpikingHAN	Heterogeneous SNN	93.7	93.3	62.9

Ablation Study

Configuration	Result	Notes
Independent GCN vs. Shared GCN	Shared is superior	Fewer parameters and reduced overfitting
IF vs. LIF vs. PLIF	PLIF is best	Learnable time constant offers greater flexibility
Timesteps T=2/4/8/16	T=4 is optimal	Too few: insufficient information; too many: wasteful computation
Hard reset vs. Subtract-threshold reset	Subtract-threshold is better	Retains more membrane potential information

Key Findings

• SpikingHAN matches PHGT on DBLP (93.7%) and closely approaches it on ACM (93.3% vs. 93.4%), demonstrating that SNNs do not sacrifice accuracy.
• The parameter count is approximately one-third that of HAN, with faster inference and significantly lower energy consumption.
• The learnable \(\tau_m\) in PLIF contributes substantially to performance — different datasets require different temporal dynamics.
• Performance on IMDB is slightly weaker (62.9% vs. 63.2%), possibly because IMDB's noisier labels cause the binary spike representations to lose some discriminative information.

Highlights & Insights

• First introduction of SNNs into heterogeneous graphs: Fills a gap in the literature and demonstrates the feasibility and efficiency advantages of this combination.
• Minimalist shared-parameter design: A single-layer shared GCN combined with semantic attention suffices to match the performance of multi-layer independent modules, embodying a "less is more" design philosophy.
• Practical value of 1-bit representations: Spike encoding compresses floating-point embeddings to 1-bit, which is highly significant for deployment on edge devices.

Limitations & Future Work

• Evaluation is limited to node classification; tasks such as link prediction and graph classification are not explored.
• Shared parameters may lack sufficient expressiveness when semantic differences across meta-paths are substantial.
• Actual energy consumption benefits on real neuromorphic hardware have not been validated.
• The performance gap on IMDB suggests that SNN binarization is more sensitive to noisy labels.

Related Work & Insights

• vs. HAN: HAN performs independent aggregation per meta-path; SpikingHAN uses shared parameters and SNNs, resulting in fewer parameters and lower energy consumption.
• vs. SpikingGCN/SpikeGCL: These methods handle only homogeneous graphs; SpikingHAN is the first to support heterogeneous graphs with multi-type nodes and edges.
• vs. PHGT: The Transformer-based approach achieves the highest accuracy but at substantial computational cost; SpikingHAN trades a marginal performance gap for significant efficiency gains.

Rating

• Novelty: ⭐⭐⭐⭐ First combination of SNNs and heterogeneous graphs
• Experimental Thoroughness: ⭐⭐⭐⭐ Three datasets, multi-dimensional efficiency analysis, and complete ablation study
• Writing Quality: ⭐⭐⭐⭐ Clear structure with intuitive comparative figures against HAN
• Value: ⭐⭐⭐⭐ Practically significant for edge deployment of heterogeneous graph learning

Related Papers

• [AAAI 2026] Kernelized Edge Attention: Addressing Semantic Attention Blurring in Temporal Graph Neural Networks
• [AAAI 2026] GT-SNT: A Linear-Time Transformer for Large-Scale Graphs via Spiking Node Tokenization
• [AAAI 2026] MUG: Meta-path-aware Universal Heterogeneous Graph Pre-Training
• [AAAI 2026] S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning
• [AAAI 2026] EchoLess: Label-Based Pre-Computation for Memory-Efficient Heterogeneous Graph Learning