使用Ray的知识图谱分布式模型研究  

A research on distributed model of knowledge graph using Ray

作　　者：方志聪 李廷元[1] Fang Zhicong;Li Tingyuan(School of Computer Science,Civil Aviation Flight Academy of China,Deyang 618300,China)

机构地区：[1]中国民用航空飞行学院计算机学院,德阳618300

出　　处：《现代计算机》2023年第5期28-32,共5页Modern Computer

摘　　要：现如今,知识图谱是众多消费者和企业应用程序的核心,其中学习到的图嵌入用于为这些应用程序的用户获取更多信息。由于知识图谱十分庞大,学习嵌入的过程是时间和资源密集型的,并且需要以分布式方式完成以利用多台机器的计算资源。通过使用Ray技术构建了一个端到端模型,用基于图神经网络的知识图嵌入模型的数据预处理和分布式训练,并将模型应用于链接预测任务,即使用知识图谱嵌入来发现图中节点之间的链接。与非分布式相比,在2个Ray通道的情况下实现了3倍的训练加速,而评估指标恶化几乎可以忽略不记。Nowadays,knowledge graph is at the core of numerous consumer and enterprise applications,andthe graph embed⁃ding learned from it are used to derive insights for the users of these applications.Since knowledge graph can be very large,the pro⁃cess of learning embedding is time and resource intensive and need to be done in a distributed manner to leverage compute re⁃sources of multiple machines.By using Ray,an end⁃to⁃end model is build,which embeds data preprocessing and distributed train⁃ing of the model with a knowledge graph based on a graph neural network.The model is applied to the link prediction task,that is,using knowledge graph embedding to discover links between nodes in a graph.Compared to non⁃distribution,and achieves a 3x training speedup with 2 Ray channels,while the evaluation metric deterioration is almost negligible.

关 键 词：知识图谱 分布式 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]