图自适应原型网络的小样本节点分类方法

Few-Shot Node Classification Method of Graph Adaptive Prototypical Networks

小样本节点分类旨在让机器从少量节点中学习到快速认知和分类的能力,现有小样本节点分类模型的分类性能容易受到图编码器提取的节点特征不够准确和子任务中支撑集实例的类内异常值的影响.为此,文中提出图自适应原型网络(Graph Adaptive Prototypical Networks,GAPN)的小样本节点分类方法.首先,将图中的节点通过图编码器嵌入度量空间中.然后,将全局重要度和局部重要度的融合结果作为支撑集实例的权重计算类原型,使查询集实例能自适应地学习更鲁棒的类原型.最后,计算自适应任务的类原型与查询集实例之间距离产生的分类概率,最小化分类概率和真实标签间的正间隔损失,反向更新网络参数,学习到更有判别性的节点特征.在常用图数据集上的实验表明,文中方法具有较优的节点分类性能.

Few-shot node classification aims to make machines recognize and classify quickly from a small number of nodes.Existing few-shot node classification models are easily affected by the inaccurate node features extracted by encoders and the intra-class outliers of query set instances in sub-tasks.Therefore,a graph adaptive prototypical networks(GAPN)model is proposed.Firstly,the nodes are embedded into the metric space by the graph encoder.Then,prototypes are computed by fusing the global importance and the local importance as weight of support set instances,and thus more robust prototypes can be learned adaptively for query set instances.Finally,the distance between the class prototypes of the adaptive task and the query set instance is calculated to generate the classification probability.By minimizing the positive marginal feedback loss between the classification probability and the true label,network parameters are updated backward and more discriminative node features can be learned.Experimental results on common graph datasets show that GAPN model yields better node classification performance.