链路预测是理解复杂网络结构与功能的关键任务。传统方法依赖人工设计的特征,难以捕捉非线性关联,而现有图神经网络(Graph Neural Network,GNN)在聚合邻居信息时往往忽视节点间关系的差异性。为此,文章提出一种融合多头注意力机制的图...链路预测是理解复杂网络结构与功能的关键任务。传统方法依赖人工设计的特征,难以捕捉非线性关联,而现有图神经网络(Graph Neural Network,GNN)在聚合邻居信息时往往忽视节点间关系的差异性。为此,文章提出一种融合多头注意力机制的图神经网络链路预测方法(Graph Neural Network with Multi-Head Attention for Link Prediction,GNN-MHA)。该模型首先利用GNN学习节点的嵌入表示,进而引入多头注意力机制,在多个特征子空间中自适应地加权邻居关键信息,深层挖掘节点间的潜在关联。在多个真实网络数据集上的实验表明,所提模型在AUC等指标上优于其他基准方法,展现了优越的链路预测性能。展开更多
知识图谱嵌入通过将实体和关系投影到连续的低维向量空间中,为机器学习模型提供更强大的知识表示输入,从而支撑更多的知识图谱应用场景。近年来,研究人员试图利用知识图谱中的本体和实例之间的潜在语义信息来增强知识图谱的嵌入。然而,...知识图谱嵌入通过将实体和关系投影到连续的低维向量空间中,为机器学习模型提供更强大的知识表示输入,从而支撑更多的知识图谱应用场景。近年来,研究人员试图利用知识图谱中的本体和实例之间的潜在语义信息来增强知识图谱的嵌入。然而,它们未能有效融合概念的层次结构和实例的特定信息,并且忽略了isA关系之间的传递性,导致模型在处理知识图谱中的长尾实体时的性能和泛化能力受限。为了弥补上述不足,提出了一个融合了本体和实例的知识图谱嵌入模型JMOI(Representation Learning of Knowledge Graph via Jointly Modeling Ontology and Instances)。该模型通过引入自注意力机制,能够捕捉到概念和实例之间复杂的语义关系,并增加了一个可学习的参数来调整概念嵌入的邻域范围,以区分不同概念的层次信息,从而对isA关系的传递性进行建模。在YAGO26K-906和DB111K-174数据集上的实验结果表明,与现有技术相比,JMOI在大多数情况下都达到了最佳性能,与次优模型相比,在链接预测Hits@1指标上最大提升了6.5%,在三元组分类中召回率指标最大提升了6.9%。展开更多
The design synthesis is the key issue in the mechanical conceptual design to generate the design candidates that meet the design requirements.This paper devotes to propose a novel and computable synthesis approach of ...The design synthesis is the key issue in the mechanical conceptual design to generate the design candidates that meet the design requirements.This paper devotes to propose a novel and computable synthesis approach of mechanisms based on graph theory and polynomial operation.The graph framework of the synthesis approach is built firstly,and it involves:(1)the kinematic function units extracted from mechanisms;(2)the kinematic link graph that transforms the synthesis problem from mechanical domain into graph domain;(3)two graph representations,i.e.,walk representation and path representation,of design candidates;(4)a weighted matrix theorem that transforms the synthesis process into polynomial operation.Then,the formulas and algorithm to the polynomial operation are presented.Based on them,the computational flowchart to the synthesis approach is summarized.A design example is used to validate and illustrate the synthesis approach in detail.The proposed synthesis approach is not only supportive to enumerate the design candidates to the conceptual design of a mechanical system exhaustively and automatically,but also helpful to make that enumeration process computable.展开更多
文摘链路预测是理解复杂网络结构与功能的关键任务。传统方法依赖人工设计的特征,难以捕捉非线性关联,而现有图神经网络(Graph Neural Network,GNN)在聚合邻居信息时往往忽视节点间关系的差异性。为此,文章提出一种融合多头注意力机制的图神经网络链路预测方法(Graph Neural Network with Multi-Head Attention for Link Prediction,GNN-MHA)。该模型首先利用GNN学习节点的嵌入表示,进而引入多头注意力机制,在多个特征子空间中自适应地加权邻居关键信息,深层挖掘节点间的潜在关联。在多个真实网络数据集上的实验表明,所提模型在AUC等指标上优于其他基准方法,展现了优越的链路预测性能。
文摘知识图谱嵌入通过将实体和关系投影到连续的低维向量空间中,为机器学习模型提供更强大的知识表示输入,从而支撑更多的知识图谱应用场景。近年来,研究人员试图利用知识图谱中的本体和实例之间的潜在语义信息来增强知识图谱的嵌入。然而,它们未能有效融合概念的层次结构和实例的特定信息,并且忽略了isA关系之间的传递性,导致模型在处理知识图谱中的长尾实体时的性能和泛化能力受限。为了弥补上述不足,提出了一个融合了本体和实例的知识图谱嵌入模型JMOI(Representation Learning of Knowledge Graph via Jointly Modeling Ontology and Instances)。该模型通过引入自注意力机制,能够捕捉到概念和实例之间复杂的语义关系,并增加了一个可学习的参数来调整概念嵌入的邻域范围,以区分不同概念的层次信息,从而对isA关系的传递性进行建模。在YAGO26K-906和DB111K-174数据集上的实验结果表明,与现有技术相比,JMOI在大多数情况下都达到了最佳性能,与次优模型相比,在链接预测Hits@1指标上最大提升了6.5%,在三元组分类中召回率指标最大提升了6.9%。
基金Supported by State Key Program of National Natural Science Foundation of China(Grant No.51535009)111 Project of China(Grant No.B13044).
文摘The design synthesis is the key issue in the mechanical conceptual design to generate the design candidates that meet the design requirements.This paper devotes to propose a novel and computable synthesis approach of mechanisms based on graph theory and polynomial operation.The graph framework of the synthesis approach is built firstly,and it involves:(1)the kinematic function units extracted from mechanisms;(2)the kinematic link graph that transforms the synthesis problem from mechanical domain into graph domain;(3)two graph representations,i.e.,walk representation and path representation,of design candidates;(4)a weighted matrix theorem that transforms the synthesis process into polynomial operation.Then,the formulas and algorithm to the polynomial operation are presented.Based on them,the computational flowchart to the synthesis approach is summarized.A design example is used to validate and illustrate the synthesis approach in detail.The proposed synthesis approach is not only supportive to enumerate the design candidates to the conceptual design of a mechanical system exhaustively and automatically,but also helpful to make that enumeration process computable.
