Research And Implementation Of Representation Learning Algorithm For Graph Structured Data

Multiple relational data in physical,biological,social,and information systems can be modeled as homogeneous or heterogeneous conceptual graph,and efficient modeling of graph can promote the development of reseach and applications on graphs.Representation learning is one of the important research directions in the field of Artificial Intelligence(AI),which plays a key role in related fields such as natural language processing and image semantic alignment.Using representation learning techniques to extract and express features of graph-structured data can better support the analysis and application of graphs.This thesis mainly studies the representation learning method of graph,which is to generate the semantic information representation of entities and relationships by projecting the entities or relationships in the graph into the low-dimensional vector space.After a thorough analysis of the design ideas and modeling mechanisms of the existing graph embedding algorithms,it is concluded that these methods have the following shortcomings: First,the current mainstream work neglects the semantic diversity of the target object.There are semantic paradoxes in the basic assumptions of these models,which lead to low semantic resolution of entities and relationships and a bottleneck in the accuracy with dealing with complex relationship types.Second,the homogenous and heterogeneous graphs cannot be uniformly modeled by the existing work,which indicates that the representation learning solutions in one specific field cannot implement application migration.Third,the symmetry/asymmetry of the relationship plays an important role in solving the semantic diversity of the relationship,but the existing work does not deeply analyze and model the symmetry/asymmetry of the relationship,and the performance of the algorithm is inevitable affected.This thesis studies the above problems and put forward some corresponding solutions.The main contributions are as follows:1.We propose a nonlinear embedding algorithm based on the analysis of the problems existing in the modeling process of the representation learning algorithm,which uses the stacked neural network to self-encode the semantics of entities and relationships.Moreover,the accuracy and recall rate of the algorithm is improved through a two-way training strategy.The experimental results on several benchmarks indicate that the proposed algorithm significantly outperforms the current mainstream related work.2.After studying and exploring the rules of association between nodes in different types of complex networks,a unified graph representation learning algorithm is proposed.We designed a knowledge learning mechanism based on Multi-shot from the perspective of human logical reasoning thinking mode.Then we proposed a generalized graph representation learning framework based on deep neural network.Experimental evidence has strongly supported the validity of the proposed modeling ideas,indicating that the knowledge representation learning framework can be effectively applied to various types of network environments.3.Considering the influence of the symmetry/asymmetry of the relationship on the performance of the algorithm,this paper proposes an algorithm design idea of relational mirroring,and proposes a self-encoding model which using a recurrent neural network to encode the semantic combination of entities and relationships,and then using deep neural networks for semantic decoding.This model effectively solves the problem of modeling and representation learning of asymmetric relationships.The experimental results on the public benchmark dataset show that the proposed algorithm outperforms the related work in the accuracy and recall rate of relational reasoning tasks and multi-label classification tasks.