Adaptive Randomized Graph Neural Network Based on Markov Diffusion Kernel

Qianli Ma, Zheng Fan, Chenzhi Wang, Yuhua Qian

Abstract: Graph neural networks (GNNs) especially Graph convolutional networks (GCNs) GCNs, are popular in graph representation learning. However, GCNs only consider the nearest neighbor, which makes it difficult to expand the node domain, and their performance decline as the number of layers increases due to the over-smoothing problem. Therefore, this paper proposes an Adaptive Randomized Graph Neural Network based on Markov Diffusion Kernel (ARM-net) to overcome these limitations. Firstly, ARM-net designs a random propagation strategy based on Bernoulli distribution; Secondly, an adaptive propagation process based on Markov diffusion kernels is designed to separate feature transformations from propagation, expand the node domain and reduce the risk of model’s over-smoothing; Finally, a graph regularization term is added to enable nodes to find more information useful for their classification results, thereby improve the generalization performance of the model. Experimental results show that the model outperforms several recently proposed semi-supervised classification algorithms in semi-supervised node classification tasks. ARM-net has better performance on multiple data sets. Through experiments, ARM-net solves the over-smoothing problem encountered during GNN propagation to some extent, and the model has better generalization performance.

External IDs:dblp:conf/icann/MaFWQ23