2023-arxiv Bake off redux-a review and experimental evaluation of recent time series classification algorithms

Motivation

• We extend the taxonomy to include three categories to reflect the recent development, which contain convolution、feature and deep learning.
• We introduce 30 new datasets based on UCR112.
• We make a summary of recent time series classification algorithms.

Method

1. The summary about recent TSC methods, which can be categorized to eight types based on representation of data:

• Distance: classification is based on some time series specific distance
  measure between series
• Feature: global features are extracted and passed to a standard classifier in a simple pipeline
• Interval: features are derived from selected phase dependent intervals
  in an ensemble of pipelines, and further output the summary statics.
• Shapelet: phase independent discriminatory subseries form the basis
  for classification(classify throught the existence of shapelet)
• Dictionary: histograms of counts of repeating patterns are the features for a classifier
• Convolution: convolutions and pooling operations create the feature
  space for classification
• Deep Learning: neural network
• Hybrid: approaches combine two or more of the above approaches

1. The summary about recent TSC methods, which can be categorized to three types based on design patterns:

• single pipeline:transformations+standard ML classifier;
• ensembles of pipelines;
• transformations embedded in classifier structure, such decision tree,where the data is transformed at each node.

note: there are generally two kinds of transformations:
(1)series-to-vector,such as calculating summary features;
(2)series-to-series,such as transforming to the frequency domain of the series;

1. The eight types contain methods:

• Distance: 1-NN DTW, Elastic Ensemble(EE), Proximityforest(PF), ShapeDTW
• Feature: The canonical time series characteristics(Catch22), Time Series Feature extraction based on Scalable Hypothesis Tests(TSFresh), Generalized signatures, FreshPRINCE
• Interval: Time Series Forest(TSF), Random interval spectral ensemble(RISE), Supervised Time Series Forest(STSF), Random STSF(RSTSF), The Canonical Interval Forest(CIF), The diverse representation canonical interval forest(DrCIF)
• Shapelet: The Shapelet Transform Classifer(STC), The Generalised Random Shapelet Forest(RSF), The Multible representation sequence learner(MrSEQL), MrSQM, Random Dilated shaplet transform(RDST)
• Dictionary: Bag-of-SFA-Symbols(BOSS), Word Extraction for Time Series Classification(WEASEL), WEASEL with dilation(WEASEL-D), Contractble BOSS(cBOSS), Spatial BOSS, The Temporal Dictionary Ensemble(TDE)
• Convolution: ROCKET, MiniRocket, MultiRocket, Hydra;
• Deep learning: InceptionT, Resnet
• Hybrid: HCa, HC1, HC2, TS-CHIEF

The detailed information can be seen as follow:

Result

• 30 new univariate datasets introduced:

• Evaluation on UCR112:

• Evaluation on UCR142: