TSFuse

Python package for automatically constructing features from multiple time series

Installation

Install the latest release using pip:

pip install tsfuse

Quickstart

The example below shows the basic usage of TSFuse.

Data format

The input of TSFuse is a dataset where each instance is a window that consists of multiple time series and a label.

Time series

Time series are represented using a dictionary where each entry represents a univariate or multivariate time series. As an example, let's create a dictionary with two univariate time series:

from pandas import DataFrame
from tsfuse.data import Collection
X = {
    "x1": Collection(DataFrame({
        "id":   [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3],
        "time": [0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2],
        "data": [1, 2, 3, 1, 2, 3, 3, 2, 1, 3, 2, 1],
    })),
    "x2": Collection(DataFrame({
        "id":   [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3],
        "time": [0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2],
        "data": [1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3],
    })),
}

The two univariate time series are named x1 and x2 and each series is represented as a Collection object. Each Collection is initialized with a DataFrame that has three columns:

• id which is the identifier of each instance, i.e., each window,
• time which contains the time stamps,
• data contains the time series data itself.

For multivariate time series data, there can be multiple columns similar to the data column. For example, the data of a tri-axial accelerometer would have three columns x, y, z instead of data as it simultaneously measures the x, y, z acceleration.

Labels

There should be one target value for each window, so we create a Series where the index contains all unique id values of the time series data and the data consists of the labels:

from pandas import Series
y = Series(index=[0, 1, 2, 3], data=[0, 0, 1, 1])

Feature construction

To construct features, TSFuse provides a construct function which takes time series data X and target data y as input, and returns a DataFrame where each column corresponds to a feature. In addition, this function can return a computation graph which contains all transformation steps required to compute the features for new data:

from tsfuse import construct
features, graph = construct(X, y, return_graph=True)

To apply this computation graph to new data, simply call transform with a time series dictionary X as input:

features = graph.transform(X)

Documentation

The documentation is available on https://arnedb.github.io/tsfuse/

Citing TSFuse

If you use TSFuse for a scientific publication, please consider citing this paper:

De Brabandere, A., Op De Beéck, T., Hendrickx, K., Meert, W., & Davis, J. TSFuse: automated feature construction for multiple time series data. Machine Learning (2022)

@article{tsfuse,
    author  = {De Brabandere, Arne
               and Op De Be{\'e}ck, Tim
               and Hendrickx, Kilian
               and Meert, Wannes
               and Davis, Jesse},
    title   = {TSFuse: automated feature construction for multiple time series data},
    journal = {Machine Learning},
    year    = {2022},
    doi     = {10.1007/s10994-021-06096-2},
    url     = {https://doi.org/10.1007/s10994-021-06096-2}
}