MrSQM: Fast Time Series Classification with Symbolic Representations

MrSQM (Multiple Representations Sequence Miner) is a time series classifier. The MrSQM method can quickly extract features from multiple symbolic representations of time series and train a linear classification model with logistic regression. The method has four variants with four different feature selection strategies:

• MrSQM-R: Random feature selection.
• MrSQM-RS: MrSQM-R with a follow-up Chi2 test to filter less important features.
• MrSQM-S: Pruning the all-subsequence feature space with a Chi2 bound and selecting the optimal set of top k subsequences.
• MrSQM-SR: Random sampling of the features from the output of MrSQM-S.

Installation

Dependencies

cython >= 0.29
numpy >= 1.18
pandas >= 1.0.3
scikit-learn >= 0.22
fftw3 (http://www.fftw.org/)

Installation using pip

pip install mrsqm

Installation from source

Download the repository:

git clone https://github.com/mlgig/mrsqm.git

Move into the code directory of the repository:

cd mrsqm/mrsqm

Build package from source using:

pip install .

Example

Load data from arff files

X_train,y_train = util.load_from_arff_to_dataframe("data/Coffee/Coffee_TRAIN.arff")
X_test,y_test = util.load_from_arff_to_dataframe("data/Coffee/Coffee_TEST.arff")

Train with MrSQM

clf = MrSQMClassifier(nsax=0, nsfa=5)
clf.fit(X_train,y_train)

Make predictions

predicted = clf.predict(X_test)

More examples can be found in the example directory, including a Jupyter Notebook with detailed steps for training, prediction and explanation. The full UEA and UCR Archive can be downloaded from http://www.timeseriesclassification.com/.

This repository provides supporting code, results and instructions for reproducing the work presented in our publication:

"Fast Time Series Classification with Random Symbolic Subsequences", Thach Le Nguyen and Georgiana Ifrim https://project.inria.fr/aaltd22/files/2022/08/AALTD22_paper_5778.pdf

"MrSQM: Fast Time Series Classification with Symbolic Representations and Efficient Sequence Mining", Thach Le Nguyen and Georgiana Ifrim https://arxiv.org/abs/2109.01036

Citation

If you use this work, please cite as:

@article{mrsqm2022,
  title={Fast Time Series Classification with Random Symbolic Subsequences},
  author={Le Nguyen, Thach and Ifrim, Georgiana},
  year={2022},
  booktitle = {AALTD},
  url = {https://project.inria.fr/aaltd22/files/2022/08/AALTD22_paper_5778.pdf},
  publisher={Springer}
}
@article{mrsqm2022-extended,
  title={MrSQM: Fast Time Series Classification with Symbolic Representations and Efficient Sequence Mining},
  author={Le Nguyen, Thach and Ifrim, Georgiana},
  year={2022},
  booktitle = {arxvi},
  url = {https://arxiv.org/abs/2109.01036},
  publisher={}
}