Knowledge Graph Representation And Application With Graph Structure

A knowledge graph(KG)is a structured knowledge base that represents entities and their relationships in the objective world in the form of a graph.By using knowledge representation and reasoning techniques,objective knowledge in the KG can be provided to artificial intelligence systems,giving them the ability to solve complex tasks similar to humans.Therefore,the KG,which describes common sense and facts,has become a widely used knowledge representation method in both academia and industry.At the same time,Graph Neural Network(GNN)has shown excellent performance in tasks such as graph representation learning,information propagation,and link prediction.Considering that the KG itself is precisely a type of graph-structured data,using a graph to construct the association between knowledge and data,and applying GNN technology,can potentially combine KG and GNN to achieve better explainable and trustworthy artificial intelligence.This dissertation analyzes existing KG representation learning and application based on GNN technology and proposes targeted improved algorithms for the problems faced by existing KG representation learning algorithms.This dissertation mainly focuses on the key issue of KG representation learning and researches the algorithm studies of GNN on traditional KGs,multimodal KGs,uncertain KGs,and how to optimize graph structure information to enhance recommendation systems.Specifically,the main research contents of this dissertation include the following:(1)For traditional KGs,this dissertation designs a novel GNN model.Existing KGs are usually incomplete due to noise problems caused by manual construction or automatic construction algorithms,resulting in non-connected graphs.This problem makes existing GNNs based on information propagation principles ineffective in KG representation,i.e.,unable to effectively capture the complete structural information in the KG.Therefore,this dissertation proposes a multi-relational GNN based on the maximum mutual information on the graph.The model borrows the idea of unsupervised contrastive learning,encoding the global and local representations of the KG by using an adaptive relational graph attention network and maximizing the mutual information between the global and local representations to learn complete structural information.At the same time,the model learns semantic information in the KG by using another adaptive relational graph attention network and combines the previous structural information to generate a complete KG embedding representation.(2)For multimodal KGs,this dissertation designs a GNN model based on low-rank tensor fusion.Most of the existing KG representation learning based on GNNs is for single-modal KGs,ignoring the fact that entities in KGs in real-life scenarios may contain multiple modal forms of information,such as images,speech,etc.To address the issue of multimodal information representation and graph structure information representation,this dissertation proposes a hyper-node relational graph attention network to achieve multimodal KG representation learning.The model consists of three parts: the information fusion module,the relational graph neural network module,and the prediction verification module.In the first step,the information fusion module fuses the pre-trained feature vectors of various modalities of entities into a multimodal feature using low-rank tensor fusion,obtaining a hyper-node representation vector for each entity.In the second step,for the hyper-graph,the model uses the adaptive relational graph attention network proposed earlier to capture the structural information in the multimodal KG and obtain the final hyper-node representation and relation representation that includes graph structure information.In the third step,to verify the obtained KG representation learning,the model uses the prediction verification module to conduct link prediction experiments on the KG.(3)For uncertain KGs,this dissertation proposes a GNN model that combines box embedding representation to address uncertain knowledge graphs.Knowledge graphs often contain noisy and erroneous fact triplets during their construction,leading to a new structured knowledge repository called uncertain knowledge graph.Each fact triplet in the uncertain knowledge graph has a confidence score to measure its probability of being true,making representation learning for uncertain knowledge graphs require considering confidence score estimation on top of entity-relation embedding.However,existing representation learning methods for uncertain knowledge graphs often ignore graph structural information.To overcome this problem,this dissertation presents a rectangular graph neural network for representation learning on uncertain knowledge graphs.The model uses a unique geometric embedding representation(i.e.,rectangular embedding representation)to replace traditional vector embedding representation and employs affine transformations to represent the relationships between entity rectangles.Additionally,this dissertation defines a novel message-passing paradigm for obtaining graph structural information in the rectangular embedding representation of uncertain knowledge graphs.(4)For application problems of KGs,this dissertation proposes a graph-based nonsampling learning method for knowledge-enhanced recommendation systems.The data in recommendation systems usually is implicit feedback data,which is typically processed using negative sampling or non-sampling strategies.However,traditional negative sampling strategies use random sampling to characterize the distribution of implicit feedback data,which cannot fully reflect the real data distribution.Based on the non-sampling strategy,this dissertation presents a graph non-sampling learning method that utilizes graph structural information,including two core techniques: graph-based non-sampling strategy and multi-hop top-K neighborhood aggregation mechanism.The graph-based nonsampling strategy considers the graph structural information of the user-item bipartite graph and the knowledge graph,and uses node centrality to assign weights to negative samples,achieving significant performance improvement compared to traditional nonsampling strategies.The multi-hop top-K neighborhood aggregation mechanism first selects the top-K important neighbor nodes with high centrality rankings around the target node,and then aggregates the information of the sampled neighbor nodes to update the representation of the target node,balancing efficiency and performance.Finally,this dissertation summarizes the research content and provides a future outlook based on feasible research ideas and directions.