MultiRocket

Multiple pooling operators and transformations for fast and effective time series classification

Preprint: arxiv:2102.00457

<div align="justify">We propose MultiRocket, a fast time series classification (TSC) algorithm that achieves state-of-the-art performance with a tiny fraction of the time and without the complex ensembling structure of many state-of-the-art methods. MultiRocket improves on MiniRocket, one of the fastest TSC algorithms to date, by adding multiple pooling operators and transformations to improve the diversity of the features generated. In addition to processing the raw input series, MultiRocket also applies first order differences to transform the original series. Convolutions are applied to both representations, and four pooling operators are applied to the convolution outputs. When benchmarked using the University of California Riverside TSC benchmark datasets, MultiRocket is significantly more accurate than MiniRocket, and competitive with the best ranked current method in terms of accuracy, HIVE-COTE 2.0, while being orders of magnitude faster.</div>

Reference

If you use any part of this work, please cite:

@article{Tan2021MultiRocket,
  title={{MultiRocket}: Multiple pooling operators and transformations for fast and effective time series classification},
  author={Tan, Chang Wei and Dempster, Angus and Bergmeir, Christoph and Webb, Geoffrey I},
  year={2021},
  journal={arxiv:2102.00457}
}

Requirements

All python packages needed are listed in requirements.txt file and can be installed simply using the pip command.

• pandas
• numpy
• numba
• sklearn
• catch22 (optional)
• sktime (required for MTSC)
• torch (required for logistic regression)

Code

The main.py file contains a simple code to run the program on a single UCR dataset.

The main_ucr_109.py file runs the program on all 109 UCR datasets.

The main_mtsc.py file contains a simple code to run the program on a single MTSC dataset.

Arguments:
-d --data_path          : path to dataset
-p --problem            : dataset name
-i --iter               : determines the resample of the UCR datasets
-n --num_features       : number of features 
-t --num_threads        : number of threads (> 0)
-s --save               : 0=don't save results, 1=save results
-v --verbose            : verbosity

Results

These are the results on 30 resamples of the 109 UCR Time Series archive from the UCR Time Series Classification Archive. MultiRocket is on average the current most accurate scalable TSC algorithm, that is not significantly less accurate than HIVE-COTE 2.0.

<p align="center"> <img src="results/figures/cd_resamples_30_top_5.png"/> </p> <p align="center"> <img src="results/figures/cd_resamples_30.png"/> </p> <p float="center" align="center"> <img src="results/figures/annotated_scatter_hc2_multirocket.png" width="400" align="center"/> <img src="results/figures/annotated_scatter_minirocket_multirocket.png" width="400" align="center"/> </p>

The following shows the total compute time for 109 UCR datasets. Compute times are averaged over 30 resamples of 109 UCR datasets and run on a cluster using 32 threads on AMD EPYC 7702 CPUs.

<p float="center" align="center"> <img src="results/figures/time_32_cores_resamples_10,50.png" width="400" align="center"/> <img src="results/figures/time_32_cores_resamples_10,50_vs_minirocket_50k.png" width="400" align="center"/> </p>

The following compares the total training time for 112 UCR datasets against some state of the arts. The first four algorithms are computed using a single thread on AMD EPYC 7702 CPU. The single thread timing results can be found in training time, test time and total time. The rest of the results are obtained from the paper HIVE-COTE 2.0: a new meta ensemble for time series classification

<div align="center">

| TSC algorithm | Total train time | |---------------|------------------| | MiniRocket (default 10k features) | 2.44 minutes | | MultiRocket (10k features) | 4.38 minutes | | MiniRocket (50k features) | 5.25 minutes | | MultiRocket (default 50k features) | 15.77 minutes | | Rocket | 2.85 hours | | Arsenal | 27.91 hours | | DrCIF | 45.40 hours | | TDE | 75.41 hours | | InceptionTime (GPU) | 86.58 hours | | STC | 115.88 hours | | HIVE-COTE 2.0 | 340.21 hours | | HIVE-COTE 1.0 | 427.18 hours | | TS-CHIEF | 1016.87 hours |

</div>

The following table contains the averaged accuracy over 30 resamples of 109 UCR datasets, found in results. The results for other classifiers can also be obtained from timeseriesclassification.com.

|dataset_name|MultiRocket|MultiRocket_100k|MultiRocket_10k|Arsenal|BOSS|CIF|DrCIF|HIVE-COTE 2.0|HIVE-COTEv1_0|InceptionTime|MiniRocket|ProximityForest|RISE|ROCKET|ResNet|S-BOSS|STC|TDE|TS-CHIEF|TSF|WEASEL|cBOSS| |------------|-----------|----------------|---------------|-------|----|---|-----|-------------|-------------|-------------|----------|---------------|----|------|------|------|---|---|--------|---|------|-----| |ACSF1|0.833666667|0.834666667|0.826666667|0.805333333|0.768333333|0.767|0.784333333|0.833333333|0.85|0.826666667|0.822333333|0.638333333|0.76|0.807|0.824|0.815|0.838333333|0.797666667|0.807|0.635|0.818|0.757333333| |Adiac|0.823103154|0.822421142|0.811167945|0.772122762|0.749019608|0.767263427|0.809633419|0.795140665|0.796248934|0.822250639|0.801705067|0.722165388|0.757971014|0.771952259|0.81543052|0.742710997|0.793179881|0.751577153|0.779710145|0.711935209|0.798806479|0.74629156| |ArrowHead|0.895238095|0.896|0.893714286|0.862666667|0.868761905|0.82|0.825333333|0.886285714|0.876|0.880380952|0.880952367|0.883619048|0.828190476|0.859047619|0.858666667|0.887809524|0.806666667|0.900571429|0.881142857|0.796761905|0.848380952|0.877904762| |BME|1|1|1|0.997555556|0.865777778|0.996888889|0.999555556|0.999555556|0.982222222|0.996444444|0.992222133|0.999111111|0.786|0.997333333|0.999111111|0.865111111|0.929777778|0.911555556|0.996444444|0.962444444|0.947777778|0.784666667| |Beef|0.77|0.773333333|0.77|0.755555556|0.612222222|0.725555556|0.791111111|0.796666667|0.735555556|0.682222222|0.7611111|0.594444444|0.742222222|0.76|0.676666667|0.655555556|0.735555556|0.683333333|0.632222222|0.688888889|0.74|0.571111111| |BeetleFly|0.896666667|0.895|0.895|0.886666667|0.943333333|0.868333333|0.878333333|0.903333333|0.963333333|0.893333333|0.906666667|0.86|0.871666667|0.885|0.853333333|0.936666667|0.933333333|0.941666667|0.958333333|0.833333333|0.886666667|0.975| |BirdChicken|0.888333333|0.886666667|0.893333333|0.886666667|0.983333333|0.866666667|0.948333333|0.951666667|0.94|0.951666667|0.908333333|0.903333333|0.868333333|0.881666667|0.945|0.968333333|0.87|0.971666667|0.963333333|0.815|0.865|0.976666667| |CBF|0.99262963|0.993|0.992555556|0.995962963|0.998925926|0.986481482|0.984777778|0.997925926|0.998259259|0.996111111|0.996222333|0.99362963|0.949037037|0.995925926|0.988222222|0.999074074|0.985296296|0.997407407|0.998444444|0.971851852|0.979777778|0.99837037| |Car|0.92|0.922777778|0.918333333|0.915555556|0.848333333|0.812222222|0.843888889|0.907777778|0.868888889|0.901111111|0.9211111|0.805555556|0.753333333|0.911666667|0.908333333|0.859444444|0.858333333|0.87|0.878888889|0.766111111|0.834444444|0.843333333| |Chinatown|0.966666667|0.966958212|0.966958212|0.967826087|0.877065112|0.963188406|0.979903382|0.973566569|0.962779397|0.964917396|0.968999|0.948019324|0.888532556|0.96686103|0.970144928|0.879105928|0.962973761|0.955458937|0.961835749|0.952964043|0.957337221|0.94763285| |ChlorineConcentration|0.780711806|0.782586806|0.762274306|0.780286458|0.658246528|0.731579861|0.739427083|0.769019097|0.733949653|0.86359375|0.753350767|0.631137153|0.764756944|0.796137153|0.841006944|0.659114583|0.735225694|0.693402778|0.660842014|0.723064236|0.754861111|0.664956597| |CinCECGTorso|0.949613527|0.947077295|0.949589372|0.865652174|0.914758454|0.987294686|0.993961353|0.998236715|0.993743961|0.832753623|0.875797133|0.937657005|0.947439614|0.864130435|0.767874396|0.94263285|0.977777778|0.98326087|0.953429952|0.958188406|0.984975845|0.829154589| |Coffee|1|1|0.997619048|1|0.985714286|0.995238095|0.998809524|0.998809524|0.992857143|0.998809524|0.998809533|0.991666667|0.98452381|1|0.996428571|0.980952381|0.989285714|0.992857143|0.99047619|0.986904762|0.989285714|0.99047619| |Computers|0.8508|0.8504|0.8412|0.842266667|0.800533333|0.7732|0.816666667|0.858533333|0.811066667|0.8656|0.793866667|0.714266667|0.778933333|0.842933333|0.8604|0.82|0.799066667|0.818133333|0.753866667|0.6488|0.778533333|0.7788| |CricketX|0.82|0.822136752|0.815213675|0.833589744|0.762393162|0.762649573|0.772735043|0.83991453|0.816153846|0.853247863|0.825128233|0.80042735|0.706239316|0.838974359|0.808119658|0.784102564|0.792051282|0.814273504|0.83042735|0.692735043|0.775726496|0.764273504| |CricketY|0.841709402|0.841367521|0.831709402|0.841452991|0.75042735|0.778376068|0.799059829|0.85042735|0.80982906|0.86|0.8413674|0.79982906|0.709145299|0.845042735|0.81017094|0.771367521|0.777948718|0.803333333|0.817008547|0.685897436|0.77957265|0.751282051| |CricketZ|0.837692308|0.838974359|0.825128205|0.852307692|0.769316239|0.781880342|0.792393162|0.859487179|0.833931624|0.861111111|0.8424787|0.802820513|0.721623932|0.853247863|0.813247863|0.786752137|0.807179487|0.828461538|0.838205128|0.705897436|0.78991453|0.772478632| |Crop