VS-Graph: Scalable and Efficient Graph Classification Using Hyperdimensional Computing

TL;DR

VS-Graph introduces a topology-centric, hyperdimensional framework for graph classification that avoids gradient-based training by using Spike Diffusion to derive topology-aware node identifiers and Associative Message Passing for multi-hop aggregation in hypervectors. The method builds a graph embedding via mean readout and performs prototype-based classification, achieving competitive accuracy with GNNs while delivering substantial training efficiency (up to 450x faster) and robustness to dimensionality reduction (down to D=128). Empirically, it outperforms prior HDC baselines by 4–5% on key datasets and maintains strong performance across multiple benchmarks, with favorable potential for edge and neuromorphic hardware deployment. The work demonstrates that a weight-free, hyperdimensional approach can close much of the gap with learnable GNNs while enabling ultra-efficient computation on resource-constrained devices.

Abstract

Graph classification is a fundamental task in domains ranging from molecular property prediction to materials design. While graph neural networks (GNNs) achieve strong performance by learning expressive representations via message passing, they incur high computational costs, limiting their scalability and deployment on resource-constrained devices. Hyperdimensional Computing (HDC), also known as Vector Symbolic Architectures (VSA), offers a lightweight, brain-inspired alternative, yet existing HDC-based graph methods typically struggle to match the predictive performance of GNNs. In this work, we propose VS-Graph, a vector-symbolic graph learning framework that narrows the gap between the efficiency of HDC and the expressive power of message passing. VS-Graph introduces a Spike Diffusion mechanism for topology-driven node identification and an Associative Message Passing scheme for multi-hop neighborhood aggregation entirely within the high-dimensional vector space. Without gradient-based optimization or backpropagation, our method achieves competitive accuracy with modern GNNs, outperforming the prior HDC baseline by 4-5% on standard benchmarks such as MUTAG and DD. It also matches or exceeds the performance of the GNN baselines on several datasets while accelerating the training by a factor of up to 450x. Furthermore, VS-Graph maintains high accuracy even with the hypervector dimensionality reduced to D=128, demonstrating robustness under aggressive dimension compression and paving the way for ultra-efficient execution on edge and neuromorphic hardware.

Figures (3)

• Figure 1: Classification Accuracy Comparison Across Benchmark Datasets. Predictive performance of VS-Graph (Ours) compared to the HDC baseline (GraphHD) and standard GNN models (GCN, GAT, GIN) across five benchmark datasets: MUTAG, PTC_FM, PROTEINS, DD, and NCI1.
• Figure 2: Effect of hypervector dimensionality on classification accuracy for VS-Graph (Ours) and GraphHD. Each subplot reports test accuracy as the hypervector size D is reduced from 8192 to 128, showing that VS-Graph remains robust under aggressive dimensionality reduction while GraphHD degrades more sharply.
• Figure 3: Training time and Inference latency per graph (in milliseconds) for VS-Graph (Ours) and GraphHD versus hypervector dimensionality.