基于子图划分的多尺度节点分类方法

MULTI-SCALE NODE CLASSIFICATION METHOD BASED ON SUBGRAPH PARTITION

为了解决深度图神经网络中存在的过平滑问题,提出一种基于子图划分的多尺度节点分类方法。该方法以Graph-Inception网络结构为核心,采用一种基于子图划分的数据预处理方法,通过改变图中的网络结构,优化特征聚集方式,有效地抑制了冗余搜索带来的过平滑问题;利用不同尺寸卷积核的组合来提取目标节点多尺度邻域的特征信息,以实现对图神经网络深度扩展的等效,一定程度上抑制了深层网络结构带来的过平滑问题。实验结果表明,该方法能够有效地抑制图神经网络中出现的过平滑问题,在基准数据集PPI、Reddit和Amazon上的分类准确率都得到了不同程度的提高。

In order to solve the problem of over-smooth in the deep graph neural networks,a multi-scale node classification method based on subgraph partition is proposed.Taking the Graph-Inception as the core,this method used a subgraph partition as a data preprocess method.The network structure in the graph was changed,the method of feature aggregation was optimized,which effectively suppressed the over-smooth caused by redundant features searching.The neighborhood feature information of the target node at multi-scale was extracted through the combination of convolution kernels of different sizes,so as to achieve the equivalent of the depth expansion of the graph neural networks.To a certain extent,the over-smooth problem caused by the deep network structure was suppressed.Experimental results show that the proposed method can effectively suppress the over-smooth problem in graph neural networks,and improve the classification accuracy on the benchmark datasets PPI,Reddit and Amazon to vary degrees.