Genetic algorithm-based robust graph convolutional network design leveraging node splitting and fusion

doi:10.1016/j.swevo.2026.102366

Genetic algorithm-based robust graph convolutional network design leveraging node splitting and fusion

2026-03-20

0

PRE

AI

DOI：10.1016/j.swevo.2026.102366

Original

Original request for help

Save

Abstract

En 中文

Neural network systems have consistently demonstrated superior performance across diverse domains, while their application in safety-critical environments continues to face significant robustness challenges due to soft errors. Traditional fault-tolerant approaches, including hardware redundancy and fault-aware training, are often insufficient and fail to exploit the intrinsic biological and evolutionary principles underlying neural networks. Inspired by neural Darwinism, this paper proposes a genetic-algorithm-based structural optimization framework for graph convolutional networks, which employs node splitting to generate redundant units and node fusion to integrate functionally similar units. Iterative network evolution progressively shifts information processing from highly centralized, sensitive nodes to more robust distributed representations. Evaluated through irradiation experiments, the proposed approach can limit fault propagation, mitigate critical errors, and reshape fault distribution patterns, all without incurring additional computational overhead. Its application in practical safety-critical scenarios provides an efficient, low-cost pathway for automatically constructing robust graph neural networks.

Keywords:

genetic algorithm

graph convolutional network

node splitting

node fusion

fault tolerance

Genetic algorithm-based robust graph convolutional network design leveraging node splitting and fusion

Abstract

Journal

Organization