pyprotoclust 0.1.0

Pyprotoclust is an implementatin of representative hierarchical clustering using minimax linkage.

The original algorithm is from Hierarchical Clustering With Prototypes via Minimax Linkage by Jacob Bien and Robert Tibshirani.

Pyprotoclust takes a distance matrix as input. It returns a linkage matrix encoding the hierachical clustering as well as an additional list labelling the prototypes associated with each clustering. This allows a user to integrate with the existing tools in the SciPy hierarchical clustering module.

Installation:

pip install pyprotoclust

Usage:

from pyprotoclust import protoclust
import numpy as np
import scipy as sp
import scipy.cluster.hierarchy
import scipy.spatial.distance

# Generate two-dimensional toy data
n = 60
np.random.seed(4)
params = [{'mean': [-7, 0], 'cov': [[1, 1], [1, 5]]},
          {'mean': [1, -1], 'cov': [[5, 0], [0, 1]]},
          {'mean': [3, 7], 'cov': [[1, 0], [0, 1]]}]
data = np.vstack([np.random.multivariate_normal(p['mean'], p['cov'], n) for p in params])
X = sp.spatial.distance.squareform(sp.spatial.distance.pdist(data))

# Produce a hierarchical clustering using minimax linkage
Z, prototypes = protoclust(X)

# Generate clusters at a set cut_height using scipy's hierarchy module
cut_height = 7
T = sp.cluster.hierarchy.fcluster(Z, cut_height, criterion='distance')
L,M = sp.cluster.hierarchy.leaders(Z, T)

# Get the prototypes associated with the generated clusters
P = data[prototypes[L]]

The previous example produces a linkage matrix Z and prototypes P that can be used to produce dendrograms and other plots of the data.