Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised Feature Based Algorithms for Time Series Extrinsic Regression

David Guijo-Rubio, Matthew Middlehurst, Guilherme Arcencio, Diego Furtado Silva, Anthony Bagnall

TL;DR

This study demonstrates that both algorithms, along with InceptionTime, exhibit significantly better performance compared to the other 18 regressors tested, and more importantly, DrCIF is the only one that significantly outperforms a standard rotation forest regressor.

Abstract

Time Series Extrinsic Regression (TSER) involves using a set of training time series to form a predictive model of a continuous response variable that is not directly related to the regressor series. The TSER archive for comparing algorithms was released in 2022 with 19 problems. We increase the size of this archive to 63 problems and reproduce the previous comparison of baseline algorithms. We then extend the comparison to include a wider range of standard regressors and the latest versions of TSER models used in the previous study. We show that none of the previously evaluated regressors can outperform a regression adaptation of a standard classifier, rotation forest. We introduce two new TSER algorithms developed from related work in time series classification. FreshPRINCE is a pipeline estimator consisting of a transform into a wide range of summary features followed by a rotation forest regressor. DrCIF is a tree ensemble that creates features from summary statistics over random intervals. Our study demonstrates that both algorithms, along with InceptionTime, exhibit significantly better performance compared to the other 18 regressors tested. More importantly, these two proposals (DrCIF and FreshPRINCE) models are the only ones that significantly outperform the standard rotation forest regressor.

Unsupervised Feature Based Algorithms for Time Series Extrinsic Regression

TL;DR

This study demonstrates that both algorithms, along with InceptionTime, exhibit significantly better performance compared to the other 18 regressors tested, and more importantly, DrCIF is the only one that significantly outperforms a standard rotation forest regressor.

Abstract

Time Series Extrinsic Regression (TSER) involves using a set of training time series to form a predictive model of a continuous response variable that is not directly related to the regressor series. The TSER archive for comparing algorithms was released in 2022 with 19 problems. We increase the size of this archive to 63 problems and reproduce the previous comparison of baseline algorithms. We then extend the comparison to include a wider range of standard regressors and the latest versions of TSER models used in the previous study. We show that none of the previously evaluated regressors can outperform a regression adaptation of a standard classifier, rotation forest. We introduce two new TSER algorithms developed from related work in time series classification. FreshPRINCE is a pipeline estimator consisting of a transform into a wide range of summary features followed by a rotation forest regressor. DrCIF is a tree ensemble that creates features from summary statistics over random intervals. Our study demonstrates that both algorithms, along with InceptionTime, exhibit significantly better performance compared to the other 18 regressors tested. More importantly, these two proposals (DrCIF and FreshPRINCE) models are the only ones that significantly outperform the standard rotation forest regressor.
Paper Structure (52 sections, 21 figures, 6 tables, 1 algorithm)

This paper contains 52 sections, 21 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and related work
  3. Time Series Classification (TSC) algorithms
  4. Unsupervised Feature-Based Regressors
  5. FreshPRINCE
  6. DrCIF
  7. Methodology
  8. Regression Algorithms
  9. Experimental Design
  10. Results
  11. Recreating Results on the 19 TSER datasets
  12. Benchmarking the new TSER archive
  13. Evaluation of new TSER algorithms
  14. Analysis
  15. Run time
  16. ...and 37 more sections

Figures (21)

  • Figure 1: Examples of soil spectrograms used to predict potassium concentration in the PotassiumConcentration dataset.
  • Figure 2: Diagrams visualising the DrCIF transformation (left) and DrCIF ensemble structure (right).
  • Figure 3: Reproduction of the RMSE ranks on the original archive (19 datasets and 5 resamples). Left is the original image from tan2021time. Right is our recreation.
  • Figure 4: RMSE ranks for 13 Regressors used in tan2021time on 63 TSER datasets.
  • Figure 5: RMSE ranks for 11 regressors on 62 TSER datasets.
  • ...and 16 more figures