Research On Temporal Knowledge Graph Completion Based On Recurrent Graph Neural Network

A knowledge graph is a visual structure of a knowledge base and plays a vital role in applications in information retrieval,natural language understanding,recommender systems,and medical care.Since the incompleteness of knowledge graphs limits its application in downstream tasks,knowledge graph completion,which is dedicated to solving the problem of missing knowledge graph data,has become a current research hotspot.However,knowledge graphs extracted,organized and constructed from real-world resources often contain facts with uneven distribution and complex temporal dynamics.Due to the fact that previous static knowledge graph completion methods are highly dependent on sufficient training examples and do not consider the temporal properties of facts,their performance in completion tasks is inevitably restricted.Therefore,this paper focuses on the temporality of facts,the long-tail problem and the sparsity of relations in knowledge graph completion,and conducts research in three aspects.(1)Aiming at the temporality of facts in knowledge graphs,this paper proposes a temporal knowledge graph completion method(TiRGN)that fuses local and global historical patterns.TiRGN utilizes the laws of historical development to comprehensively consider the sequence,repetition and circulation patterns of historical facts in the knowledge graph,and balance the local and global historical features of events.(2)Aiming at the long-tailed distribution of relations within a single moment in temporal knowledge graphs,this paper proposes a few-shot knowledge graph completion method(StarRing)based on matching networks.StarRing leverages an adaptive multi-aggregator to learn two types of neighborhood topologies for a given relational entity pair,combining coarse-grained semantic matching with fine-grained relevance matching.(3)Aiming at the sparsity of relations in temporal knowledge graphs at multiple times,this paper proposes a few-shot temporal knowledge graph completion method(RAN)based on relational adaptive networks.RAN utilizes a relational adaptive network to dynamically adjust local history lengths for different low-frequency and highfrequency relations,and perform dynamic path matching for few-shot relations.This paper validates the proposed method on public datasets.The performance of TiRGN is improved on all 6 public datasets,and the experimental results demonstrate the effectiveness of capturing multiple historical patterns.StarRing is validated on three few-shot knowledge graph completion datasets from NELL,FB15k237 and Wiki.The experimental results show that using two types of topology to model a more complete neighborhood structure effectively alleviates the long tail problem.After synthesizing the first two methods of RAN for simultaneous multi-task learning for entity prediction and relation prediction,RAN conducts experiments on multi-relational datasets.Experimental results show that it outperforms previous methods on both entity prediction and relation prediction tasks in temporal knowledge graph completion,and its performance improvement is more significant on few-shot relations.