EdgeGNN: Efficient GNN Architecture Search for Resource-Constrained Devices

• 1、Beijing Jiaotong University, School of Computer Science and Technology, Beijing,100044

Abstract：Graph Neural Networks (GNNs) excel at processing non-Euclidean data, but their design often necessitates substantial knowledge and iterative adjustments. While Neural Architecture Search (NAS) has successfully automated the design of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), applying it to GNNs, particularly focusing on optimization for resource-constrained devices, remains a challenge. This paper aims to reduce the dependence on specialized knowledge and iterative adjustments in GNN architecture design, with particular emphasis on efficient deployment on resource-constrained devices. To achieve this, we propose a reinforcement learning-based neural architecture search framework, EdgeGNN, that automatically discovers high-accuracy, low-latency GNN architectures. Our approach employs a joint optimization strategy, taking into account both overall model accuracy and inference latency influenced by the hardware. Experimental results on multiple benchmark datasets demonstrate that the GNN architectures identified by our framework significantly outperform existing baselines in terms of model size and inference speed, while maintaining competitive predictive accuracy.

Keywords： Artificial Intelligence Graph Neural Network Neural Architecture Search