基于GraphSage节点度重要性聚合的网络节点分类研究

Network Node Classification Based on GraphSage Node Degree Importance Aggregation

传统的图嵌入算法及图神经网络模型在对网络节点分类时仅使用了节点本身的属性信息或者特征信息,很少使用节点在网络中的结构信息。如何在图神经网络聚合时引入节点网络结构信息来提升分类准确性也是一个值得深入研究的问题。因此,在GraphSage模型的基础上,根据网络中节点度及节点重要性设计了新的聚合函数并提出了GraphSage-Degree模型。首先,模型根据节点度获得节点在邻域中的重要性,然后再以重要性为依据来聚合节点的特征,使得网络中重要的节点能够尽可能的聚合更多的特征信息,并且在GraphSage-Degree中设置了一个与节点度有关的超参数D,能够通过调节该参数D使得在不同的数据集上达到最佳分类状态。在Cora、Citeseer和Pubmed 3个公开数据集上进行了测试,GraphSage-Degree与其他方法相比,macro-F1的平均提升值分别为8.72%、10.37%和8.29%,在Pubmed上有最大提升值38.84%;micro-F1的平均提升值分别为8.97%、11.16%和6.9%,在Pubmed上有最大提升值38.39%。

The traditional Graph Embedding algorithm and graph neural network model only use the attribute information or feature information of the nodes themselves when classifying the nodes in the network,but rarely use the structural information of the nodes in the network.It is also a problem worthy of further study to introduce the node network structure information to improve the classification accuracy when aggregating the graph neural network.Therefore,based on the GraphSage model,a new aggregation function was designed and a new GraphSage-Degree model was proposed on the node degree and node importance in the network.Firstly,the model obtained the importance of the nodes in the neighbourhood based on the node degree,then aggregated the features of the nodes relied on the importance,so that the important nodes in the network can aggregate as much feature information as possible,and set a hyperparameter D related to the node degree in the GraphSage-Degree,which can be adjusted to achieve the best classification state on different data sets.The GraphSage-Degree was tested on three publicly available datasets,Cora,Citeseer and Pubmed.The average improvement values of macro-F1 compared respectively with other methods are 8.72%,10.37%and 8.29%,with the maximum improvement value of 38.84%on Pubmed.The average improvement values of micro-F1 are 8.97%,11.16%,and 6.9%,respectively,with a maximum boost value of 38.39%on Pubmed.