Multi-kernel Times Series Outlier Detection.

Time series are sequences of observations ordered by time. Detecting outliers in a set of time series is very important for many use cases, including fraud detection and predictive maintenance. However, this task continues to be difficult: First, time series may be of different lengths and conventional distance measures like the Euclidean distance can not capture their similarity well. Workarounds like feature engineering require domain knowledge and render solutions domain-specific. Second, many existing techniques are supervised, but training labels are expensive if not impossible to obtain. In this paper, we propose Multi-Kernel Times Series Outlier Detection (MK-TSOD), a method that combines the Fourier Transform, Global Alignment Kernels, and Multiple Kernel Learning with Support Vector Data Description. We describe its specifics, and show that MK-TSOD outperforms existing methods on standard benchmark data.