基于属性关联及匹配差异度的数据流异常检测

Outlier Detection Model for Data Streams Based on Attribute Associations and Match Difference Degree

为解决类别属性数据流异常点检测问题,针对事务数据流环境,提出了基于属性关联及匹配差异度的数据流异常检测模型AAMDD(attribute associations and match difference degree).AAMDD模型离线构建一个关联规则库,并对其进行增量式更新.同时,利用时间敏感型滑动窗口(time-sensitive sliding windows,TimeSW)维护数据流数据,每经过一个时间跨度,就将当前窗口中每条数据包含的项集与关联规则库进行匹配,计算匹配差异度,根据匹配差异度的不同在线检测异常点.此外,给出了与AAMDD模型相对应的算法AAMDD-algorithm.实验结果表明,AAMDD-algorithm比FODFP-Stream算法的效率和检测精确度分别平均提高了37.43%和5.51%,并且AAMDD-algorithm的查全率保持在77%以上,可用于事务型数据流异常检测.

In order to solve the problem of outlier detection for categorical data streams, an outlier detection model for data streams based on attribute associations and match difference degree was proposed, called as AAMDD. This model builds an association rule library off-line and updates it with the incremental method. Meanwhile, it maintains the data streams by using time-sensitive sliding windows （TimeSW）. In a time step, the AAMDD matches data in current window with association rules in the association rule library and calculates the match difference degree （MDD）. Then, oufliers can be identified on-line through different MDDs. An algorithm for the AAMDD was given, called as AAMDD-algorithm. The experiment results show that compared with the FODFP-Stream algorithm, the AAMDD-algorithm has on average 5.51% and 37.43% improvements respectively in detection precision and efficiency, and its recall is above 77%. It can be used to detect outliers in transaction data streams.