Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SPROCKET: Extending ROCKET to Distance-Based Time-Series Transformations With Prototypes

Nicholas Harner

TL;DR

SPROCKET extends ROCKET by injecting distance-based, prototype-driven features into the convolutional transform framework for time series classification. By selecting a small, logarithmically growing set of prototypes and a careful mix of distance measures, it creates distance-derived features that complement ROCKET kernels. Across 98 UCR/UEA datasets, SPROCKET substantially boosts ensemble performance when combined with HYDRA and MultiROCKET, with the MR-HY-SP ensemble achieving top average rankings; however, elastic distances incur higher compute, prompting a practical Euclidean variant. The work highlights a promising direction that bridges convolutional kernel methods and distance-based classification, while outlining concrete avenues for optimization and broader empirical testing.

Abstract

Classical Time Series Classification algorithms are dominated by feature engineering strategies. One of the most prominent of these transforms is ROCKET, which achieves strong performance through random kernel features. We introduce SPROCKET (Selected Prototype Random Convolutional Kernel Transform), which implements a new feature engineering strategy based on prototypes. On a majority of the UCR and UEA Time Series Classification archives, SPROCKET achieves performance comparable to existing convolutional algorithms and the new MR-HY-SP ( MultiROCKET-HYDRA-SPROCKET) ensemble's average accuracy ranking exceeds HYDRA-MR, the previous best convolutional ensemble's performance. These experimental results demonstrate that prototype-based feature transformation can enhance both accuracy and robustness in time series classification.

SPROCKET: Extending ROCKET to Distance-Based Time-Series Transformations With Prototypes

TL;DR

SPROCKET extends ROCKET by injecting distance-based, prototype-driven features into the convolutional transform framework for time series classification. By selecting a small, logarithmically growing set of prototypes and a careful mix of distance measures, it creates distance-derived features that complement ROCKET kernels. Across 98 UCR/UEA datasets, SPROCKET substantially boosts ensemble performance when combined with HYDRA and MultiROCKET, with the MR-HY-SP ensemble achieving top average rankings; however, elastic distances incur higher compute, prompting a practical Euclidean variant. The work highlights a promising direction that bridges convolutional kernel methods and distance-based classification, while outlining concrete avenues for optimization and broader empirical testing.

Abstract

Classical Time Series Classification algorithms are dominated by feature engineering strategies. One of the most prominent of these transforms is ROCKET, which achieves strong performance through random kernel features. We introduce SPROCKET (Selected Prototype Random Convolutional Kernel Transform), which implements a new feature engineering strategy based on prototypes. On a majority of the UCR and UEA Time Series Classification archives, SPROCKET achieves performance comparable to existing convolutional algorithms and the new MR-HY-SP ( MultiROCKET-HYDRA-SPROCKET) ensemble's average accuracy ranking exceeds HYDRA-MR, the previous best convolutional ensemble's performance. These experimental results demonstrate that prototype-based feature transformation can enhance both accuracy and robustness in time series classification.

Paper Structure

This paper contains 32 sections, 5 equations, 39 figures, 7 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Intuition
  3. ROCKET Architecture and Modification
  4. Number of Prototypes
  5. Prototype Selection
  6. Distance Measures
  7. Algorithm Analysis
  8. Experimental Results
  9. Settings
  10. Prototype and Distance Selection
  11. Ensemble Experiments
  12. Kernel Scaling
  13. Large Scale MSM Benchmarking
  14. Large Scale Euclidean Benchmarking
  15. Future Work
  16. ...and 17 more sections

Figures (39)

  • Figure 1: Relative comparisons of Distance Measures with the SPROCKET Transform.
  • Figure 2: Stratified Results maintain a similar rank ordering to true random.
  • Figure 3: Total Feature Transformation Time For Each Distance, Random Prototypes.
  • Figure 4: Ensemble Grid for Random Prototypes.
  • Figure 5: Ensemble Grid for Stratified Random Prototypes.
  • ...and 34 more figures