API Reference / algorithms / timeseries /

classification/

Time-series classification.

Classifying a whole series by its shape, rather than forecasting its next value. Input is a panel of shape (n_samples, n_channels, n_timepoints); a 2-D array is read as univariate.

This is a different task from everything in the parent timeseries package, which forecasts. It is also different from ordinary classification: flattening a series into columns and handing it to a feature-matrix classifier discards the time ordering that carries the signal.

Algorithms

MiniRocketClassifier: 84 fixed dilated kernels summarised by proportion

of positive values, then a linear head. Near state-of-the-art accuracy at a tiny fraction of the cost, and the right default at any real scale.

HIVECOTEClassifier: Combines all of the below, weighting each by its own

cross-validated accuracy. The most accurate and by far the most expensive.

TimeSeriesForestClassifier: Mean, standard deviation and slope of random

intervals, in a forest. The temporally localised option: it says where in the series the difference lives.

BOSSClassifier: Turns each series into a bag of symbolic words built

from low-frequency Fourier coefficients. The noise-tolerant option: the low-pass step discards exactly the detail noise lives in.

ShapeletTransformClassifier: Finds the short subsequences that separate

the classes, and represents each series by its distance to them. The interpretable option: the fitted shapelets are real subsequences you can plot and read.

DTWNeighborsClassifier: Nearest neighbour under Dynamic Time Warping.

The standard baseline of the field, and still hard to beat — but its cost grows with the training set, which MINIROCKET's does not.

Transforms

MiniRocketTransformer: The MINIROCKET features without a classifier, for

pipelines and non-classification uses.

Distances

dtw_distance / dtw_pairwise: Elastic alignment that tolerates stretching

and compression of the time axis.

lb_keogh / lb_keogh_envelope: The cheap lower bound that makes

nearest-neighbour search under DTW affordable.

Notes

Z-normalise each series before using an elastic distance unless the absolute level is genuinely meaningful — otherwise DTW largely measures offset. A Sakoe-Chiba band of roughly 10% of the series length is the usual starting point: it speeds the computation up and typically improves accuracy by forbidding degenerate alignments.