Abstract:

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.

Key words: anomaly detection, time series, transfer learning