TR2023-091

GPU-APUMPEDI: A Parallel Algorithm for Computing Approximate Pan Matrix Profiles of Time Series


    •  Zhang, J., Nikovski, D., Nakamura, T., "GPU-APUMPEDI: A Parallel Algorithm for Computing Approximate Pan Matrix Profiles of Time Series", International conference on Time Series and Forecasting, July 2023.
      BibTeX TR2023-091 PDF
      • @inproceedings{Zhang2023jul2,
      • author = {Zhang, Jing and Nikovski, Daniel and Nakamura, Takaaki},
      • title = {GPU-APUMPEDI: A Parallel Algorithm for Computing Approximate Pan Matrix Profiles of Time Series},
      • booktitle = {International conference on Time Series and Forecasting},
      • year = 2023,
      • month = jul,
      • url = {https://www.merl.com/publications/TR2023-091}
      • }
  • MERL Contact:
  • Research Areas:

    Data Analytics, Machine Learning

Abstract:

The Matrix Profile (MP) of a test time series with respect to a base time series has been shown to be a versatile primitive for many data mining tasks including time series anomaly detection. The MP records distances from all subsequences in the test time series to their respective nearest neighbors in the base time series. The Pan Ma- trix Profile (PMP) is a matrix with each row being an MP corresponding to a single subsequence length, and computing explicitly an exact PMP is slow. We propose a GPU-based approximation algorithm called GPU- APUMPEDI to compute the PMP under unnormalized Euclidean dis- tance by combining GPU-based MP algorithms and linear interpolation. We validate its efficiency and effectiveness through extensive numerical experiments on the UCR Anomaly Archive.