Research On Link Prediction Based On Knowledge Graph Representation Learning

Incomplete data sources and immature knowledge extraction technologies lead to incomplete knowledge graphs which are constructed manually or automatically.Knowledge graph link prediction aims to infer new knowledge based on existing knowledge,so as to achieve knowledge graph completion.This paper focuses on link prediction of large-scale open domain knowledge graphs.In the existing link prediction algorithms,the high time complexity of the path ranking algorithm will cause a large time overhead when applied to large-scale knowledge graphs.Therefore,this paper proposes a path ranking algorithm based on dual-sampling random walks,which aims to improve the operation efficiency of the algorithm.Based on this,considering that translation-based link prediction algorithms ignore multi-step relation paths and complex inference patterns,this paper proposes an improved algorithm that integrates multi-step paths generated by dual-sampling random walks into translation-based algorithms,in order to improve the accuracy of link prediction.The main research contents of this paper are as follows:(1)A dual-sampling path ranking algorithm(DSPRA)is proposed.This paper improves the sampling strategy of the path ranking algorithm,and proposes a random walk method based on the double-layer sampling mechanism.The sampling mechanism perform particle sampling in the relation layer and the entity layer respectively,which reduces the time complexity of the random walk-based path ranking algorithm without significantly affecting prediction accuracy.The improvement of efficiency lays a foundation for reducing the time overhead of the subsequent fusion algorithm.(2)A TransE algorithm incorporating dual sampling path constraints(DSP-TransE)is proposed.Path confidence is proposed and multi-step paths mined by dual-sampling random walks are merged into TransE according to a translation hypothesis incorporating path constraints,so that the TransE model can simultaneously use direct relations and multi-step relation paths to perform representation learning.The algorithm thus obtains more accurate knowledge representation vectors and improves the accuracy of link prediction.At the same time,under the influence of the high efficiency of the DSPRA algorithm,the increase in the time overhead of the fusion algorithm is significantly reduced.Link prediction experiments are conducted on the large-scale open domain knowledge graphs,which verifies the effectiveness of the DSPRA algorithm and the DSP-TransE algorithm.