基于实例的迁移时间序列异常检测算法研究

Research on Transfer Time Series Anomaly Detection Algorithm Based on Instance

时间序列异常检测不管在学术界还是工业界都正引起人们极大的兴趣,但同时也存在异常标签数据缺失严重这一普遍问题。为了解决该问题,文章提出了基于实例的迁移时间序列异常检测算法——InsTransAnomalyDetect算法。该算法通过构建有效的决策函数来迁移实例,将原来的无监督异常检测任务转化为监督学习的任务。算法集成两种决策函数,分别是基于密度的决策函数和基于聚类的决策函数。文章最后将该方法与两种经典的异常检测算法在24个数据集上进行效果对比。实验结果表明,在24个数据集中,文中算法的表现优于无监督的异常检测算法的数据集有21个,平均准确率提升20%左右。实验证明了文中算法的优越性。

Time series anomaly detection is attracting great interest both in academia and industry. A common and ubiquitous problem is the lack of abnormal tag data. In order to solve this problem, this paper proposes InsTransAnomalyDetect which is a time series anomaly detection algorithm based on transfer learning. This algorithm transforms the instance by constructing an effective transfer decision function. According to the decision function, this algorithm transforms the original unsupervised anomaly detection task into a supervised learning task. The algorithm integrates two decision functions, which are density-based decision function and cluster-based decision function. Finally, the method is compared with two classical anomaly detection algorithms on 24 data sets. The experimental results show that 21 of the 24 data sets outperform the unsupervised anomaly detection algorithm, and the average accuracy rate is about 20% better. Experiments show that the migration learning method is promising and proves the superiority of the algorithm framework.