milia.cluster.models.tskmeans#

Time-series k-means with Euclidean, DTW or soft-DTW distance (tslearn).

Classes

TSKMeans([n_clusters, metric, n_init, ...])

k-means over whole sequences, each row of X being one series.

TSKMeansConfig(*[, name, kind, space, ...])

Validated hyperparameters of the time-series k-means clusterer.

class milia.cluster.models.tskmeans.TSKMeans(n_clusters=3, metric='euclidean', n_init=3, max_iter=20, random_state=0, band=None)#

Bases: BaseClusterer

k-means over whole sequences, each row of X being one series.

metric="dtw" aligns profiles that share a shape but are shifted in time, which plain k-means on the same rows would split.

Parameters:
  • n_clusters (int) – Number of clusters.

  • metric (Literal['euclidean', 'dtw', 'softdtw']) – "euclidean", "dtw" or "softdtw".

  • n_init (int) – Restarts.

  • max_iter (int) – Iterations per run.

  • random_state (int | None) – Seed.

  • band (int | None) – Sakoe-Chiba radius in readings for metric="dtw"; the alignment is unconstrained when omitted.

Examples

>>> import numpy as np
>>> t = np.linspace(0, 2 * np.pi, 24)
>>> X = np.vstack([np.sin(t + s) for s in (0.0, 0.1, 0.2)] + [np.cos(3 * t + s) for s in (0.0, 0.1, 0.2)])
>>> TSKMeans(n_clusters=2).fit(X).labels_.tolist()
[1, 1, 1, 0, 0, 0]
class milia.cluster.models.tskmeans.TSKMeansConfig(*, name='tskmeans', kind='clusterer', space=SearchSpace(n_clusters=(2, 12), metric=['euclidean', 'dtw']), n_clusters=3, metric='euclidean', n_init=3, max_iter=20, random_state=0, band=None)#

Bases: BaseModelConfig

Validated hyperparameters of the time-series k-means clusterer.

Fields mirror TSKMeans’s constructor; space is the default Optuna search space. Build the clusterer with make("clusterer", "tskmeans", ...) or TSKMeans(**config.params()).

Examples

>>> TSKMeansConfig().name
'tskmeans'