Parallel And Distributed Knowledge Graph Embedding

Knowledge graph is the knowledge system in which people represent and construct knowledge into structured information,which has attracted more and more attentions in recent years.As the most important application of knowledge graph,reasoning on knowledge graph has attracted many explorations through the years.Recently,knowl-edge graph embedding methods are proposed to tackle the task of reasoning on knowl-edge graph.Due to its promising performance,knowledge graph embedding methods have become the mainstream of knowledge graph reasoning.The basic idea of knowl-edge graph embedding is to embed the entities and relations of knowledge graph into low-dimensional vectors,based on which reasoning is completed with simple linear computations.Although knowledge graph embedding methods have achieved promising accuracy on reasoning tasks,all existing methods are all serial(single-thread)methods.When deal-ing with large scale knowledge graphs,these methods are too slow or even cannot be finished.To address this problem,this article designs innovative parallel and distribut-ed knowledge graph embedding methods for efficient knowledge graph embedding.These methods successfully improve the eficiency of knowledge graph embedding.The main innovative contributions are listed as follows:1.We design a new multi-thread based parallel knowledge graph embedding frame-work called ParaGraphE,which is the first parallel framework for knowledge graph embedding in the world.Based on this framework,we design many parallel knowl-edge graph embedding methods.The experimental results on real-world datasets show that the proposed knowledge graph embedding methods can sufficiently improve the speed of knowledge graph embedding without influencing the accuracy.2.We propose a distributed knowledge graph embedding method based on item-based parameter partitioning,called IBP,which is the first distributed knowledge graph em-bedding method in the world.For IBP,we design a distributed knowledge graph em-bedding method with the parameter server framework.The experimental results on real-world datasets show that IBP can significantly improve the efficiency of the origi-nal single-thread methods.3.We propose a distributed knowledge graph embedding method based on dimension-based parameter partitioning,called DBP.In order to solve the communication cost problem in IBP,we propose a distributed framework which can dramatically reduce the communication cost among multiple machines.The experimental results on real-world datasets show that compared to IBP,DBP can promote the computing efficiency further more based on this new framework.