Core Concepts
Two recently proposed algorithms, Hydra+MultiROCKET and HIVE-COTEv2, significantly outperform other time series classification approaches on both the current and new datasets.
Abstract
The article presents a comprehensive review and experimental evaluation of recent time series classification (TSC) algorithms. It extends a previous "bake off" study conducted in 2017, which compared 18 TSC algorithms on 85 datasets from the University of California, Riverside (UCR) archive.
The authors first summarize the core terminology and experimental procedure used in the study. They then describe the latest TSC algorithms, categorizing them into seven main types: distance-based, feature-based, interval-based, shapelet-based, dictionary-based, convolution-based, and deep learning-based. For each category, the authors review the best-performing algorithms and compare their performance on the current 112 equal-length problems in the UCR archive.
The authors also contribute 30 new univariate datasets to the TSC archive and evaluate the best-performing algorithm from each category on this expanded dataset. The results show that two recently proposed algorithms, Hydra+MultiROCKET and HIVE-COTEv2, significantly outperform other approaches on both the current and new TSC problems.
The authors analyze the factors that drive the performance of different algorithm types and discuss the merits of various approaches. They conclude by highlighting the key developments in the field and discussing future research directions.
Stats
The UCR archive has expanded from 85 to 112 equal-length datasets since the previous bake off.
The authors contributed 30 new univariate datasets to the TSC archive.
The best-performing algorithms achieve over 86% test accuracy on average across the 112 UCR datasets.
Quotes
"Two recently proposed algorithms, Hydra+MultiROCKET Dempster et al. (2022) and HIVE-COTEv2 Middlehurst et al. (2021), perform significantly better than other approaches on both the current and new TSC problems."
"The growth in popularity of TSC open source toolkits such as aeon and tslearn have made comparison and reproduction easier."