二值化身份感知图卷积神经网络  

Binary identify-aware graph convolutional network

作　　者：苏树智 卢彦丰 SU Shuzhi;LU Yanfeng(School of Computer Science and Engineering,Anhui University of Science&Technology,Huainan,Anhui 232001,China;Institute of Artificial Intelligence,Hefei Comprehensive National Science Center,Hefei,Anhui 230088,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽准南232001 [2]合肥综合性国家科学中心人工智能研究院,安徽合肥230088

出　　处：《光电子．激光》2022年第12期1280-1286,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61806006);中国博士后科学基金(2019M660149);安徽省重点研发计划国际科技合作专项(202004b 11020029);安徽高校协同创新项目(GXXT-2021-006);合肥综合性国家科学中心能源研究院项目(19KZS203)资助项目。

摘　　要：针对有限的内存资源导致图神经网络(graph neural network, GNN)无法完全加载属性图的问题,文中提出了二值化身份感知图卷积神经网络(binary identify-aware graph convolutional network, BID-GCN)。该网络通过在消息传递过程中递归地考虑节点的信息,为了获得一个给定的节点的嵌入,BID-GCN将提取以该节点为中心的自我网络,并进行多轮的异构消息传递,在自我网络的中心节点上应用与其他节点不同的参数。在消息传递过程中,对网络参数和输入节点特征进行二值化,并将原始的矩阵乘法修改为二值化以加速运算。通过理论分析和实验评估,BID-GCN可以减少网络参数和输入数据的平均约36倍的内存消耗,并加快引文网络上平均约49倍的推理速度,可以提供与全精度基线相当的性能,较好地解决内存资源有限的问题。To solve the problem that graph neural network(GNN) cannot fully load the entire attributed graph due to limited memory resources, the binary identify-aware graph convolutional network(BID-GCN) is proposed.In this network, the nodes information is considered recursively during message passing, and then in order to obtain an embedding of a given node, the BID-GCN will extract the ego network centered at that node and perform multiple rounds of heterogeneous message passing, applying different parameters to the central node of the ego network to the rest of the nodes.In this process, the network parameters and input node features are binary by the network.In addition, the original matrix multiplication is modified to be binary to speed up the operation.Through theoretical analysis and experimental evaluation, BID-GCN can reduce the memory consumption by the average approximate 36 times of both the network parameters and input data, and accelerate the inference speed by the average approximate 49 times on the citation networks.It can provide comparable performance to full precision baselines, and can better tackle the problem of limited memory resources.

关 键 词：深度学习 图卷积神经网络(GCN) 消息传递 二值化方法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]