Irregular message passing networks

作者：

Highlights：

• Irregular message passing is systematically presented for the first time.

• Eigen-subspace analysis of GNNs is first given.

• The structural defect of the graph attention mechanism is systematically presented.

• The positive impact of randomness on neural learning systems is fully explored.

• A unified geometric view of message propagation is proposed.

摘要

•Irregular message passing is systematically presented for the first time.•Eigen-subspace analysis of GNNs is first given.•The structural defect of the graph attention mechanism is systematically presented.•The positive impact of randomness on neural learning systems is fully explored.•A unified geometric view of message propagation is proposed.

论文关键词：Graph neural networks,Message passing,Power iteration,Random propagation,Random attention

论文评审过程：Received 27 February 2022, Revised 15 September 2022, Accepted 16 September 2022, Available online 22 September 2022, Version of Record 12 October 2022.

论文官网地址：https://doi.org/10.1016/j.knosys.2022.109919