Research On Temporal Knowledge Graph Reasoning Based On Variational Auto-Encoding And Diffusion Probability Distribution

Temporal knowledge graph reasoning aims to fill in missing links or facts in knowledge graphs,where each fact is associated with a specific timestamp.With the rapid development of artificial intelligence technology,temporal knowledge graph reasoning plays an increasingly important role in recommendation systems,question-answering systems,prediction systems,and many other fields,and has become a current research hotspot.Currently,most reasoning methods focus on reasoning based on past timestamps to complete missing facts,and there is little research exploring the reasoning of future facts.Predicting future facts faces the main challenge of how to effectively model historical facts,capture the evolution process of entities over time,and make accurate predictions for the future,which is more challenging.However,most existing methods have a common limitation: they focus on learning the dynamic representation of the temporal knowledge graph,and rarely consider the static features that remain unchanged over time.To address the above problems,this paper mainly does two main works as follows:(1)Temporal Knowledge Graph Reasoning based on Variational Autoencoder.This paper introduces two independent variables,dynamic latent variables and static latent variables,which respectively represent the dynamic and static features of all historical facts that occurred at past timestamps.To address the uncertainty of entity evolution over time,this paper models the dynamic and static representations of entities using a inference network based on Gaussian distribution,where the mean of the Gaussian distribution represents the expected position of the entity at the current timestamp,and the variance represents the time uncertainty.First,a inference network based on Bi-GRU is used to learn the static latent representation of historical facts,and a inference network based on non-linear discrete time transformation is used to learn the dynamic latent representation.Then,this paper uses reparameterization tricks to sample the latent variables multiple times,and combines the dynamic and static representations learned from the historical sequence into a unified representation to predict future facts.This paper conducts extensive experiments on four public datasets(Yago11k,Wikidata12 k,ICEWS14,and ICEWS05-15),and the results demonstrate the superiority of this paper’s method.(2)Temporal Knowledge Graph Reasoning based on Diffusion Probability Distribution.The previously proposed method based on variational autoencoder is prone to overfitting during training and cannot accurately capture the semantic evolution of entities.Therefore,this paper further explores on the basis of previous work.This paper establishes a bi-directional iterative process that divides the entity semantic modeling process into multiple sub-modules.Each sub-module consists of a forward noise injection transformation and a backward Gaussian sampling,responsible for modeling a small semantic evolution process of the entity.Compared to the variational autoencoder-based method,this method jointly models the explicit learning of the dynamic representation of entity semantics over time in the metric space through multiple sub-modules,which can achieve more accurate modeling.This paper conducts extensive experiments on four public datasets(Yago11k,Wikidata12 k,ICEWS14,and ICEWS05-15),and the results demonstrate the superiority of this method.