CiliaGraph: Enabling Expression-enhanced Hyper-Dimensional Computation in Ultra-Lightweight and One-Shot Graph Classification on Edge

TL;DR

CiliaGraph addresses the challenge of efficient graph classification on edge devices by blending Hyperdimensional Computing with a novel encoding and weighted aggregation strategy. The method preserves node distance isomorphism through non-uniform dynamic encoding, leverages a hyper-weight matrix that jointly encodes similarity and nodal degree, and employs one-shot learning to concatenate node attributes with topological messages, producing robust graph representations with minimal computation. Empirical results across multiple datasets show substantial memory and time savings (average ~292x memory, up to ~2341x in some cases; ~103x faster training) while achieving accuracy competitive with SOTA GNNs, particularly excelling when edge attributes are limited. The work also provides theoretical insights into low-dimensional hypervector spaces under quasi-orthogonality and practical guidelines for selecting dimensions and quantization levels in HDC-based graph classifiers, highlighting CiliaGraph’s practicality for edge deployment.

Abstract

Graph Neural Networks (GNNs) are computationally demanding and inefficient when applied to graph classification tasks in resource-constrained edge scenarios due to their inherent process, involving multiple rounds of forward and backward propagation. As a lightweight alternative, Hyper-Dimensional Computing (HDC), which leverages high-dimensional vectors for data encoding and processing, offers a more efficient solution by addressing computational bottleneck. However, current HDC methods primarily focus on static graphs and neglect to effectively capture node attributes and structural information, which leads to poor accuracy. In this work, we propose CiliaGraph, an enhanced expressive yet ultra-lightweight HDC model for graph classification. This model introduces a novel node encoding strategy that preserves relative distance isomorphism for accurate node connection representation. In addition, node distances are utilized as edge weights for information aggregation, and the encoded node attributes and structural information are concatenated to obtain a comprehensive graph representation. Furthermore, we explore the relationship between orthogonality and dimensionality to reduce the dimensions, thereby further enhancing computational efficiency. Compared to the SOTA GNNs, extensive experiments show that CiliaGraph reduces memory usage and accelerates training speed by an average of 292 times(up to 2341 times) and 103 times(up to 313 times) respectively while maintaining comparable accuracy.

Figures (8)

• Figure 1: Accuracy vs. training cost
• Figure 2: Limitations of SG-HDC and our solution
• Figure 3: Variable attribute distributions drive individual dynamic encoding for these attributes.
• Figure 4: The processes of weight involved aggregation and one-shot learning: symmetric aggregation with weights, followed by concatenation to preserve comprehensive features.
• Figure 5: Optimal quantization levels.