Research On Knowledge Repre-Sentation Learning Algorithm On Temporal Knowledge Graph

Knowledge graph is a semantic network technology that uses structured graph data to model entities,concepts,attributes and the relationship between them in the real world.Mining the hidden information in knowledge graphs is a hot research topic in related fields,and knowledge representation learning is a key research direction in the field of knowledge graphs.At present,most of the research objects are still limited to static knowledge graphs.However,in reality,most of the related systems evolve dynamically in time series,and the evolution characteristics of knowledge graphs in time cannot be ignored.Researching temporal knowledge graph methods suitable for dynamic related systems has great potential.Significance.The existing time information direct embedding method focuses on the change law of individual entities themselves,and uses time information as an additional feature dimension for representation learning,which lacks the grasp of the overall evolution characteristics of the time series knowledge map;the time information indirect embedding method focuses on the time series change law of the knowledge map as a whole,the representation effect of its overall time series change is better,but the defect is that the representation of time information and specific knowledge is separated,and the representation accuracy of specific entities is insufficient.In addition,most existing algorithms cannot distinguish the contribution of network neighborhood nodes and the contribution of knowledge graph snapshots to prediction performance at different times,resulting in insufficient focus in the learning process and insufficient ability to learn and mine time information.This paper conducts research on the following content for the time series knowledge graph representation learning algorithm:(1)Aiming at the problem that the existing temporal knowledge graph representation learning cannot take into account the overall temporal evolution characteristics and specific entity representation,a knowledge graph representation learning TKRL algorithm that combines direct embedding and indirect embedding of time information is proposed.The algorithm fully considers the extraction of the time series evolution features of individual entities and the overall time series evolution features of the time series knowledge map,uses matrix transformation and feature fusion to introduce the time series features of individual entities,and uses the recurrent neural network to introduce the overall time series features of the knowledge map.Large entities are corrected to increase timing representation accuracy.(2)Aiming at the problem that the learning process of the existing representation learning methods is not focused,an attention mechanismbased sequential knowledge representation learning enhancement model is proposed.The model fully considers the difference in the contribution of different neighborhood entities and knowledge map snapshots at different historical moments to the performance of representation learning.It uses the self-attention mechanism to distinguish the contribution of knowledge map space and time neighborhood entities,and focuses on the learning process to improve the effect.