API Reference / algorithms / clustering /

distance/

Distance metrics for measuring similarity between data points.

The metric decides what "similar" means, so it shapes a clusterer's output as much as the algorithm does. Each metric comes in two forms: a scalar function over two points, and a vectorised *_pairwise over whole matrices.

Metrics

  • euclidean_distance: Straight-line distance. The default, and what
k-means assumes.
  • manhattan_distance: Sum of absolute differences. Less sensitive to
outliers than Euclidean, and natural for grid-like data.
  • chebyshev_distance: The largest single-dimension difference.
  • minkowski_distance: The family the three above belong to, parametrised
by p (1 gives Manhattan, 2 Euclidean, ∞ Chebyshev).
  • cosine_distance: Angle between vectors, ignoring magnitude. Use for
text and other high-dimensional sparse data, where vector length reflects document size rather than content.

Helpers

  • get_distance_function: Look a metric up by name.
  • pairwise_distances / cdist / pdist: Full distance matrices.
  • DISTANCE_FUNCTIONS / PAIRWISE_FUNCTIONS: The name-to-function tables.

Notes

Every metric except cosine is scale-sensitive: a column measured in thousands overwhelms one measured in units, so scale features first.
python 
>>> import numpy as np
>>> from tuiml.algorithms.clustering.distance import euclidean_distance
>>> float(euclidean_distance(np.array([0.0, 0.0]), np.array([3.0, 4.0])))
5.0

Modules

chebyshev

Chebyshev (L-infinity) distance function.

0 classes · 2 functions
cosine

Cosine distance function.

0 classes · 2 functions
euclidean

Euclidean (L2) distance function.

0 classes · 2 functions
manhattan

Manhattan (L1) distance function.

0 classes · 2 functions
minkowski

Minkowski distance function.

0 classes · 2 functions