Distributed Framework For Knowledge Graph Representation Learning

To better complete and build large-scale knowledge graph,this paper mainly studies how to use distributed technology to implement stable and effective representation learning of massive knowledge graph.When using distributed technology for knowledge graph representation learning,it will face the challenge of stable and efficient processing of big data and large models on the premise of reducing semantic loss.So,it is essential to solve the problem by designing efficient distributed parallel mode,parameter interaction mode and effective model aggregation method.This paper proposes a distributed framework DTrans X based on decentralized hybrid parallel for knowledge graph representation learning.First of all,this paper proposes a distributed model based on hybrid parallel,which divides large-scale knowledge graph data and large-scale representation learning model,and trains small-scale data blocks and model blocks on different distributed computing units,so as to improve the efficiency of knowledge representation model training.Secondly,this paper designs a decentralized parameter training mode,in which each computing node carries out data training and model maintenance at the same time,and the distributed algorithm combines the two-layer parameter interaction architecture to interact the model vector efficiently.Finally,to reduce the semantic loss in parallel processing,this paper proposes a distributed model merging method based on the frequency weight of entity or relation words.This method upgrades the primary model to the advanced model at the same time as unifying the model copies.Experiments show that DTrans X can reduce the loss of semantics on the standard data set,overcome the generalization problem of single machine algorithm,improve the accuracy of knowledge representation on individual values,and show good acceleration effect and excellent stability on large-scale data set.Therefore,DTrans X not only ensures the quality of knowledge representation but also improves the training efficiency of representation learning model,which introduces a new solution for accurate and efficient representation learning of knowledge graph.