摘要： 为解决已有图表示学习方法复杂性较高的问题，提出一种能在维持图特征表达力的同时提升学习效率的方法。通过在神经网络表示模型中设置适当的双曲几何结构捕获图数据的基本属性，利用贝叶斯个性化排序目标最大化节点之间正确链接和错误链接的差距从而自动学习相似性信息，在所设计的神经排序模型中使用双曲距离函数计算节点之间的层次距离。在此基础上，基于黎曼梯度下降法学习节点的特征向量。实验结果表明，相对DNGR、HARP等方法，该方法能够高效地学习节点特征，而且能获得更加紧凑、更具表达力的特征向量表示。

关键词: 图表示学习, 双曲几何, 双曲面模型, 神经网络, 贝叶斯个性化排序

Abstract: To address the high complexity of existing graph representation learning methods,this paper proposes a new graph representation learning method to improve the learning efficiency while maintaining the representation performance of graph features.The method captures the basic properties of graph data by establishing appropriate hyperbolic geometry structure in the neural network representation model.Then the Bayesian Personalized Ranking(BPR) target is used to maximize the gap between the correct links and the wrong links to automatically learn the similarity information.Moreover,the hyperbolic distance function is used to calculate the hierarchical distance between the nodes in the designed neural ranking model.Finally,the model uses the Riemannian gradient descent method to learn the feature vector of nodes.Experimental results show that the proposed method can efficiently learn node features,and can provide more compact and more expressive feature vector representations than DNGR,HARP and other methods.

Key words: graph representation learning, hyperbolic geometry, hyperboloid model, neural network, Bayesian Personalized Ranking(BPR)

中图分类号: 

• TP18