基于边缘特征的图神经网络架构搜索

Graph Neural Architecture Search Based on Edge Features

针对图神经网络架构搜索方法存在注重实体特征信息、忽略边缘特征潜在信息的问题,提出一种边缘特征图神经网络架构搜索方法,采用随机梯度下降的搜索策略优化图神经网络架构以提高搜索性能.首先,设计基于边缘特征的搜索空间,用于实体和边缘特征的提取;其次,通过边的权重进行决策,逐步对边进行裁剪,得到一个无需权重共享的网络架构.在三维点云数据集ModelNet上的实验结果表明:所提方法能有效提取图神经网络的边缘信息;与连续贪婪神经网络搜索、基于资源平衡的架构搜索、含注入噪声的可微分神经网络搜索等自动搜索的网络架构和人工设计的网络架构相比,所提方法能搜索出高效的网络架构,在三维点云模型分类中具有较高的准确率.

Graph neural architecture search algorithms over-emphasize entity features and ignore latent relation features concealed in edges.The proposed method for edge features graph neural architecture search uses stochastic gradient descent to optimize the architecture and improve the search performance.Firstly,a search space based on edge features is designed for extracting entity and edge features.Secondly,the decision is made by the weight of edges and trimmed gradually to obtain a network architecture without weight sharing.Extensive experiments on ModelNet dataset show that the proposed method can effectively extract the edge information of the graph neural networks.Compared with SGAS(sequential greedy architecture search),RBNAS(resource balance neural architecture search),NoisyDARTS(noisy differentiable architecture search),and manually designed network architectures,the proposed method can find an efficient architecture and achieve a good performance in three-dimensional point cloud classification.