基于图的情境离群点检测方法

针对异常模式挖掘中的情境离群点检测问题,提出一种基于图的检测方法.首先对数据实例构建一个实例图,然后采用一个滑动窗口穿越数据实例,对处于滑动窗口内的数据实例,计算结点之间的闵可夫斯基距离作为边权值,然后采用最小生成树聚类算法对实例图进行聚类,再采用第二个滑动窗口穿越数据实例,根据窗口内的数据实例是否属于主趋势聚类赋予不同的离群值评分,不属于主趋势聚类的数据实例被认为是潜在的离群点.仿真实验和实际数据分析表明该方法在一元序列数据检测中是切实可行的,该方法具有较好的适用性和扩展性.

An Approach for Contextual Outlier Detection Based on Graph

Aiming at the contextual outlier detection problem in abnormal pattern mining, a graph-based detection method was presented, first a graph was built to represent the data instances, then a sliding window was used across the data instance, the Minkowski distance was calculated between nodes for the data instances in the sliding window, the Minkowski distance was adopted as the edges weight. Then the minimum spanning tree clustering algorithm was adopted to cluster the instances graph, finally the second sliding window was adopted cross the data instance, different outlier score was given to the data instance according to the data instance belonged to the main trends clustering or not. The data instances which did not belong to the main trends within the window clustering were considered potential outliers. Simulation experiments and real data analysis show that this algorithm is feasible in binary sequence data; and this method has good applicability and scalability.