Breaking Computational Barriers to Perform Time Series Pattern Mining at Scale and at the Edge
- Author(s): Zimmerman, Zachary Pierce
- Advisor(s): Brisk, Philip
- Keogh, Eamonn
- et al.
Uncovering repeated behavior in time series is an important problem in many domains such as medicine, geophysics, meteorology, and many more. With the continuing surge of smart/embedded devices generating time series data, there is an ever growing need to perform analysis on datasets of increasing size. Additionally, there is an increasing need for analysis at low power edge devices due to latency problems inherent to the speed of light and the sheer amount of data being recorded. The matrix profile has proven to be a tool highly suitable for pattern mining in time series; however, a naive approach to computing the matrix profile makes it impossible to use effectively in both the cloud and at the edge. This dissertation shows how, through the use of GPUs and machine learning, the matrix profile is computed more feasibly, both at cloud-scale and at sensor-scale. In addition, it illustrates why both of these types of computation are important and what new insights they can provide to practitioners working with time series data.