Results with new classifiers on more datasets now available

We have recently updated the result and would ask these be used in subsequent publications. Legacy results from the bake off are still available, see below. The UCR Archive currently contains 128 datasets. 15 of these are unequal length and one (Fungi) has a single instance per class in the train files. We have evaluated state-of-the-art classifiers on the remaining 112 problems, for 30 resamples.

Download all accuracy results and summary statistics.

Download all predictions for all folds accuracy results.

Critical difference diagram for 14 classifiers included in the new evaluation. Solid bars indicate cliques, within which there is no significant difference in rank. Tests are performed with the sign rank test using the Holm correction. Top clique of four classifiers represent the state of the art in Spring 2020.

Dictionary Based

• Bag of Symbolic-Fourier Approximation Symbols (BOSS)   link to paper   link to results   link to tsml code (Weka)   link to aeon code (Python)
• Contract BOSS (cBOSS)   link to paper   link to results   tsml   aeon
• Spatial BOSS (S-BOSS)   link to paper   link to results   tsml
• Word Extraction for Time Series classification (WEASEL)   link to paper   link to results   tsml aeon

Interval and Spectral Based

• Time Series Forest (TSF)   link to paper   link to results   tsml   aeon
• Random Interval Spectral Ensemble (RISE)   link to paper   link to results   tsml   aeon

Shapelet Based

• The Shapelet Transform Classifier (STC)   link to paper   link to results   tsml   aeon

Distance Based

• Proximity Forest (PF)   link to paper   link to results   tsml  

Deep Learning Based

• ResNet   link to paper   link to results   aeon
• InceptionTime   link to paper   link to results   aeon

Hybrids

• The Hierarchical Vote Collective of Transformation-based Ensembles (HIVE-COTE) v1.0   link to paper   link to results   tsml   aeon
• Time Series Combination of Heterogeneous and Integrated Embeddings Forest(TS-CHIEF)   link to paper   link to results   tsml  
• Random Convolutional Kernel Transform(ROCKET)   link to paper   link to results   aeon
• Canonical Time-series Characteristics (Catch22)   link to paper   link to results  

Legacy results used in the paper the great time series classification bake off are below. These are on the 85 datasets that were in the UCR archive at the time.

To download train/test results, click here

To download results averaged over 100 resamples, click here

If you want results for all individual folds, click here

Choose a tab to view the repositories results