Time series classification and clustering with Reservoir Computing

Figure 1: Overview of the RC classifier.

This library allows to quickly implement different architectures for time series data based on Reservoir Computing (RC), the family of approaches popularized in machine learning by Echo State Networks. This library is primarly design to perform classification and clustering of both univariate and multivariate time series. However, it can also be used to perform time series forecasting.

🚀 Getting Started - 📚 Documentation - 📊 Advanced examples

Installation

The recommended installation is with pip:

pip install reservoir-computing

Alternatively, you can install the library from source:

git clone https://github.com/FilippoMB/Time-series-classification-and-clustering-with-Reservoir-Computing.git
cd Time-series-classification-and-clustering-with-Reservoir-Computing
pip install -e .

Quick start

The following scripts provide minimalistic examples that illustrate how to use the library for different tasks.

To run them, download the project and cd to the root folder:

git clone https://github.com/FilippoMB/Time-series-classification-and-clustering-with-Reservoir-Computing.git
cd Time-series-classification-and-clustering-with-Reservoir-Computing

Classification

python examples/classification_example.py

You can also view the notebook or

Clustering

python examples/clustering_example.py

You can also view the notebook or

Forecasting

python examples/forecasting_example.py

You can also view the notebook or

Overview of the framework

In the following, we present the three main functionalities of this library.

Classification

Referring to Figure 1, the RC classifier consists of four different modules.

1. The reservoir module specifies the reservoir configuration (e.g., bidirectional, leaky neurons, circle topology). Given a multivariate time series $\mathbf{X}$ it generates a sequence of the same length of Reservoir states $\mathbf{H}$.
2. The dimensionality reduction module (optionally) applies a dimensionality reduction on the sequence of the reservoir's states $\mathbf{H}$ generating a new sequence $\mathbf{\bar H}$.
3. The representation generates a vector $\mathbf{r}_\mathbf{X}$ from the sequence of reservoir's states, which represents in vector form the original time series $\mathbf{X}$.
4. The readout module is a classifier that maps the representation $\mathbf{r}_\mathbf{X}$ into the class label $\mathbf{y}$, associated with the time series $\mathbf{X}$.

This library implements also the reservoir model space, a very powerful representation $\mathbf{r}_\mathbf{X}$ for the time series. Details about the methodology are found in the original paper.

The class RC_model contained in modules.py permits to specify, train and test an RC-model. Several options are available to customize the RC model, by selecting different configurations for each module.

The training and test function requires in input training and test data, which must be provided as multidimensional numpy arrays of shape [N,T,V], with:

• N = number of samples
• T = number of time steps in each sample
• V = number of variables in each sample

Training and test labels (Ytr and Yte) must be provided in one-hot encoding format, i.e. a matrix [N,C], where C is the number of classes.

from reservoir_computing.modules import RC_model

clf = RC_model()
clf.fit(Xtr, Ytr) # Training
Yhat = clf.predict(Xte) # Prediction

Clustering

The representation $\mathbf{r}_\mathbf{X}$ obtained from the representation module (step 3) can be used to perform time series clustering. The same class RC_model used for classification can be configured to directly return the time series representations, which can be used in unsupervised tasks such as clustering and dimensionality reduction.

As in the case of classification, the data must be provided as multidimensional NumPy arrays of shape [N,T,V]

from reservoir_computing.modules import RC_model

clst = RC_model(readout_type=None)
clst.fit(X)
rX = clst.input_repr # representations of the input data

The representations rX can be used to perfrom clustering using traditional clustering algorithms for vectorial data, such as those here.

Forecasting

The sequences $\mathbf{H}$ and $\mathbf{\bar H}$ obtained at steps 1 and 2 can be directly used to forecast the future values of the time series.

The class RC_forecaster contained in modules.py permits to specify, train and test an RC-model for time series forecasting.

from reservoir_computing.modules import RC_forecaster

fcst = RC_forecaster()
fcst.fit(Xtr, Ytr) # Training
Yhat = fcst.predict(Xte) # Predictions

Here, Xtr, Ytr are current and future values, respectively, used for training.

Advanced examples

The following notebooks illustrate more advanced use-cases.

• Perform dimensionality reduction, cluster analysis, and visualize the results: view or
• Probabilistic forecasting with advanced regression models as readout: view or
• Use advanced classifiers as readout: view or

Datasets

There are several datasets available to perform time series classification/clustering and forecasting.

Classification and clustering

from reservoir_computing.datasets import ClfLoader

downloader = ClfLoader()
downloader.available_datasets(details=True)  # Print available datasets
Xtr, Ytr, Xte, Yte = downloader.get_data('Libras')  # Download dataset and return data

Forecasting

Real-world time series

from reservoir_computing.datasets import PredLoader

downloader = PredLoader()
downloader.available_datasets(details=False)  # Print available datasets
X = downloader.get_data('CDR')  # Download dataset and return data

Synthetic time series

from reservoir_computing.datasets import SynthLoader

synth = SynthLoader()
synth.available_datasets()  # Print available datasets
Xs = synth.get_data('Lorenz')  # Generate synthetic time series

Citation

Please, consider citing the original paper if you are using this library in your reasearch

@article{bianchi2020reservoir,
  title={Reservoir computing approaches for representation and classification of multivariate time series},
  author={Bianchi, Filippo Maria and Scardapane, Simone and L{\o}kse, Sigurd and Jenssen, Robert},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2020},
  publisher={IEEE}
}

License

The code is released under the MIT License. See the attached LICENSE file.