图异常检测旨在从属性网络中检测出异常节点,其由于在许多应用领域如金融、电子贸易、垃圾邮件发送者检测中有着深远的实际意义而备受重视。传统的非深度学习方法只能捕捉图的浅层结构,对此,研究者们提出了基于深度神经网络的异常检测...图异常检测旨在从属性网络中检测出异常节点,其由于在许多应用领域如金融、电子贸易、垃圾邮件发送者检测中有着深远的实际意义而备受重视。传统的非深度学习方法只能捕捉图的浅层结构,对此,研究者们提出了基于深度神经网络的异常检测模型。然而,这些模型没有考虑到图中节点的中心性差异,这种差异在捕获节点的局部信息时会导致信息缺失或引入远端节点的噪声。此外,它们忽略了属性空间的特征信息,这些信息可以提供额外的异常监督信号。为此,从无监督的视角出发,提出了一种新颖的基于个性化PageRank和对比学习的图异常检测框架PC-GAD(Personalized PageRank and Contrastive Learning based Graph Anomaly Detection)。首先,提出一种动态采样策略,即通过计算图中每个节点的个性化PageRank向量确定其相应的子图采样数目,避免局部信息的缺失和引噪;其次,针对每个节点,分别从拓扑结构和属性空间的角度出发捕获节点的异常监督信号,并设计相应的对比学习目标,从而全面地学习潜在的异常模式;最后,经过多轮对比预测,根据输出的异常值得分评估每个节点的异常程度。为验证所提模型的有效性,分别在6个真实数据集上与基准模型开展了大量对比实验。实验结果验证了PC-GAD能够全面地识别出图中的异常节点,AUC值相比现有模型提升了1.42%。展开更多
Hypergraphs can accurately capture complex higher-order relationships,but it is challenging to identify their important nodes.In this paper,an improved PageRank(ImPageRank)algorithm is designed to identify important n...Hypergraphs can accurately capture complex higher-order relationships,but it is challenging to identify their important nodes.In this paper,an improved PageRank(ImPageRank)algorithm is designed to identify important nodes in a directed hypergraph.The algorithm introduces the Jaccard similarity of directed hypergraphs.By comparing the numbers of common neighbors between nodes with the total number of their neighbors,the Jaccard similarity measure takes into account the similarity between nodes that are not directly connected,and can reflect the potential correlation between nodes.An improved susceptible–infected(SI)model in directed hypergraph is proposed,which considers nonlinear propagation mode and more realistic propagation mechanism.In addition,some important node evaluation methods are transferred from undirected hypergraphs and applied to directed hypergraphs.Finally,the ImPageRank algorithm is used to evaluate the performance of the SI model,network robustness and monotonicity.Simulations of real networks demonstrate the excellent performance of the proposed algorithm and provide a powerful framework for identifying important nodes in directed hypergraphs.展开更多
本文针对多重线性PageRank问题,结合松弛技术,提出了一般形式的张量分裂迭代算法,并给出了相应的收敛性分析。进一步,结合Anderson加速技术,提出了新的张量分裂算法。In this paper, combining relaxation techniques, a general form o...本文针对多重线性PageRank问题,结合松弛技术,提出了一般形式的张量分裂迭代算法,并给出了相应的收敛性分析。进一步,结合Anderson加速技术,提出了新的张量分裂算法。In this paper, combining relaxation techniques, a general form of tensor splitting iterative algorithm is proposed for the multilinear PageRank problem, and the corresponding convergence analysis is given. Furthermore, a new tensor splitting algorithm is proposed by incorporating Anderson acceleration techniques.展开更多
文摘图异常检测旨在从属性网络中检测出异常节点,其由于在许多应用领域如金融、电子贸易、垃圾邮件发送者检测中有着深远的实际意义而备受重视。传统的非深度学习方法只能捕捉图的浅层结构,对此,研究者们提出了基于深度神经网络的异常检测模型。然而,这些模型没有考虑到图中节点的中心性差异,这种差异在捕获节点的局部信息时会导致信息缺失或引入远端节点的噪声。此外,它们忽略了属性空间的特征信息,这些信息可以提供额外的异常监督信号。为此,从无监督的视角出发,提出了一种新颖的基于个性化PageRank和对比学习的图异常检测框架PC-GAD(Personalized PageRank and Contrastive Learning based Graph Anomaly Detection)。首先,提出一种动态采样策略,即通过计算图中每个节点的个性化PageRank向量确定其相应的子图采样数目,避免局部信息的缺失和引噪;其次,针对每个节点,分别从拓扑结构和属性空间的角度出发捕获节点的异常监督信号,并设计相应的对比学习目标,从而全面地学习潜在的异常模式;最后,经过多轮对比预测,根据输出的异常值得分评估每个节点的异常程度。为验证所提模型的有效性,分别在6个真实数据集上与基准模型开展了大量对比实验。实验结果验证了PC-GAD能够全面地识别出图中的异常节点,AUC值相比现有模型提升了1.42%。
基金Project supported by the National Natural Science Foundation of China(Grant No.62166010)the Guangxi Natural Science Foundation(Grant No.2023GXNSFAA026087).
文摘Hypergraphs can accurately capture complex higher-order relationships,but it is challenging to identify their important nodes.In this paper,an improved PageRank(ImPageRank)algorithm is designed to identify important nodes in a directed hypergraph.The algorithm introduces the Jaccard similarity of directed hypergraphs.By comparing the numbers of common neighbors between nodes with the total number of their neighbors,the Jaccard similarity measure takes into account the similarity between nodes that are not directly connected,and can reflect the potential correlation between nodes.An improved susceptible–infected(SI)model in directed hypergraph is proposed,which considers nonlinear propagation mode and more realistic propagation mechanism.In addition,some important node evaluation methods are transferred from undirected hypergraphs and applied to directed hypergraphs.Finally,the ImPageRank algorithm is used to evaluate the performance of the SI model,network robustness and monotonicity.Simulations of real networks demonstrate the excellent performance of the proposed algorithm and provide a powerful framework for identifying important nodes in directed hypergraphs.
文摘本文针对多重线性PageRank问题,结合松弛技术,提出了一般形式的张量分裂迭代算法,并给出了相应的收敛性分析。进一步,结合Anderson加速技术,提出了新的张量分裂算法。In this paper, combining relaxation techniques, a general form of tensor splitting iterative algorithm is proposed for the multilinear PageRank problem, and the corresponding convergence analysis is given. Furthermore, a new tensor splitting algorithm is proposed by incorporating Anderson acceleration techniques.
