tselect 1.0.0

Package for selecting the relevant and non-redundant channels for multivariate time series classification.

TSelect

Installation

Option 1: Pip install

TSelect can be installed with pip.

pip install tselect

Option 2: Clone repository

Alternatively, the repository can be cloned with:

git clone https://github.com/ML-KULeuven/TSelect.git

Afterward, the requirements should be installed:

pip install -r requirements.txt

Known issues

On Windows, the installation of the pycatch22 package can fail. Installing the package with the following command usually fixes this.

pip install pycatch22==0.4.2 --use-deprecated=legacy-resolver

Quick start

TSelect is a package for selecting relevant and non-redundant channels from multivariate time series data (n instances, t timepoints, d channels). It accepts the following data formats as input:

• MultiIndex Pandas DataFrame (with index levels: (n, t) and d columns)
• 3D NumPy array (with shape: (n, d, t))
• a Dictionary with TSFuse Collection objects (see https://github.com/arnedb/tsfuse for more information)

The general set-up is as follows:

from tselect.channel_selectors.tselect import TSelect

# Load your data, split in train and test set, etc.
x_train, x_test = ... 
y_train, y_test = ...

channel_selector = TSelect(irrelevant_percentage_to_keep=0.6,
                           redundant_correlation_threshold=0.7)
channel_selector.fit(x_train, y_train)
x_train_selected = channel_selector.transform(x_train)
x_test_selected = channel_selector.transform(x_test)

clf = <some MTSC classifier> # Can be any classifier for multivariate time series classification
clf.fit(x_train_selected, y_train)
y_pred = clf.predict(x_test_selected)

Hyperparameters

TSelect has several hyperparameters that can be adapted to the specific dataset and use case.

The hyperparameters to configure the irrelevant channel selector:

• irrelevant_selector: bool, default=True
  • Whether to use the irrelevant channel selector.
• irrelevant_percentage_to_keep: float, default=0.6
  • The percentage of channels that are expected to be relevant. TSelect will keep this percentage of channels after the irrelevant channel selector step.
  • A value between 0 and 1, where 1 means all channels are kept.
• irrelevant_hard_threshold: float, default=0.5
  • All channels with an evaluation metric (e.g. ROCAUC) below this threshold are considered worse than random and are removed, unless this would remove all channels.

The hyperparameters to configure the redundant channel selector:

• redundant_selector: bool, default=True
  • Whether to use the redundant channel selector.
• redundant_correlation_threshold: float, default=0.7
  • The correlation threshold to use for the redundant channel selector step. Channels that make predictions with a correlation higher than this threshold are considered redundant.
  • A value between 0 and 1, where 1 means that the predictions have to be identical.

Other hyperparameters:

• validation_size: float, default=None
  • The size of the validation set used to compute the evaluation metric. If None, the validation size is derived from max(100, 0.25*nb_instances). The train set then includes the remaining instances.
• random_state: int, default=0
  • The random state to use for reproducibility.