Anomalydetection
Last updated
Was this helpful?
Last updated
Was this helpful?
##structural inconsistencies-2016
Hu, Renjun, et al. "An embedding approach to anomaly detection." Data Engineering (ICDE), 2016 IEEE 32nd International Conference on. IEEE, 2016.
our goal is to discover structural inconsistencies, identify such nodes which bring together diverse portions of the network, i.e., the anomalous nodes that connect to a number of diverse influential communities
现有的graph embedding方法都不是为了异常点识别设计的,没有考虑识别inconsistencies的关键:局部连接和社区结构,因此,作者设计了一个新的embedding方法,专门为了结构异常点的检测。
结点i的向量表示$X_i$是一个d维的向量,其中每一维$x_i^k$代表结点i和社区k的关联。模型中的社区不受限与现实中的社区。embedding的目标:
目标函数:
计算分数:
为了消除噪声的影响,对于NB(j)中的每个邻居,the entries in Xj whose values fall below the average one of Xj are regarded as non-influential and are replaced with 0.
为了将有影响力的社区和无影响力的区分开,对于$y_i^k < \theta y_i^*$的替换成0,theta是0和1之间的参数。
AScore(i) > thre,结点i是异常点。thre反映了有影响力的社区的diversity,取决于网络中社区结构的强度,也就是inner-community和inter-community edges的比例。强社区结构的网络有更小的thre。thre在[1.5, 10]之间。