Knowledge Graph Entity Alignment Via Graph Representation Learning

Knowledge graphs are capable of describing the concepts,entities,and relationships in the objective world in a structured form,and are widely applied in various fields,such as natural language processing,search engines,intelligent recommendation systems,robotics,and speech recognition.In recent years,as a large number of knowledge graphs with different types have emerged,the demand for integrating multi-source knowledge graphs has grown stronger,giving rise to entity alignment technology.Entity alignment techniques can align equivalent entities in multi-source knowledge graphs,connecting different knowledge graphs like bridges.Consequently,many researchers have recently focused on the field of entity alignment,propelling significant progress in this area.By analyzing existing entity alignment methods,this paper identifies four core prob-lems that still need to be addressed in the field of entity alignment:(1)weak encoding capabilities?(2)low scalability?(3)lack of effective decoding algorithms? and(4)lack of interpretability.Centered around these four core issues,this paper proposes multiple inno-vative mechanisms in the aspects of encoding network architecture,training loss function,decoding strategy,and iterative strategy.The main contributions of this paper are sum-marized as follows:To enhance the model’s encoding capability,this paper delves into the design flaws of previous graph neural network models,according to the task characteristics of knowledge graphs and entity alignment,proposes four innovations:(1)meta-relation representation and relation enhancement?(2)relation-aware self-attention mechanism?(3)relational re-flection transformation? and(4)bidirectional nearest neighbor iterative strategy.These four innovations significantly improve the graph neural network’s representation capabil-ity for knowledge graphs.Experimental results on public datasets demonstrate that the proposed method outperforms baseline approaches,with improvements in metrics such as Hits@1 and M RR exceeding 15%.To improve the model’s scalability,this paper optimizes the network architecture,alignment loss function,and iterative strategy in three directions.In terms of network ar-chitecture,this paper introduces a dual attention matching network that further enhances the accuracy of entity alignment methods without increasing computational complexity.In terms of alignment loss function,this paper proposes two novel entity alignment loss functions: normalized hard sample mining loss and negative-free loss,both of which sig-nificantly accelerate model training speed,with the latter also greatly reducing memory requirements.In terms of iterative strategy,this paper introduces an incremental semi-supervised learning strategy that substantially reduces model training time by only requir-ing the model to review a minimal number of past samples when learning newly generated semi-supervised samples.By combining these innovations,the training speed of the graph neural network model is increased by more than 25 times,and memory consumption on medium-sized datasets is reduced to 3 GB.In terms of decoding algorithms,this paper proposes an entity alignment decoding al-gorithm based on third-order tensor isomorphism.Specifically,this paper derives two sets of tensor isomorphism equations based on the third-order tensor isomorphism hypothesis:(1)adjacency tensor isomorphism equations and(2)Gram tensor isomorphism equations.By combining these two sets of equations,the proposed decoding algorithm can effec-tively exploit the adjacency isomorphism and internal isomorphism of knowledge graphs.Experimental results on public datasets show that the decoding algorithm brings a signif-icant performance improvement of over 3% even when applied to state-of-the-art entity alignment methods,with the additional time cost being less than 3 seconds.In terms of interpretability,this paper introduces a three-view label propagation mech-anism,enabling label propagation algorithms designed for homogeneous graphs to effec-tively operate on knowledge graphs.Furthermore,this paper proposes two approxima-tion computation strategies,random orthogonal labels and sparse Sinkhorn iteration,to reduce computational complexity and enhance scalability.The three-view label propaga-tion mechanism possesses remarkable efficiency and interpretability,allowing for a clear explanation of how entities are aligned by tracing the propagation process at each step.All the proposed methods of this paper have been peer-reviewed and published in international conferences.We hope that this work will provide valuable inspiration for further research and exploration of entity alignment problems.