Tutorial for scipy.cluster.hierarchy

Question

I’m trying to understand how to manipulate a hierarchy cluster but the documentation is too … technical?… and I can’t understand how it works.

Is there any tutorial that can help me to start with, explaining step by step some simple tasks?

Let’s say I have the following data set:

a = np.array([[0,   0  ],
              [1,   0  ],
              [0,   1  ],
              [1,   1  ], 
              [0.5, 0  ],
              [0,   0.5],
              [0.5, 0.5],
              [2,   2  ],
              [2,   3  ],
              [3,   2  ],
              [3,   3  ]])

I can easily do the hierarchy cluster and plot the dendrogram:

z = linkage(a)
d = dendrogram(z)

Now, how I can recover a specific cluster? Let’s say the one with elements [0,1,2,4,5,6] in the dendrogram?
How I can get back the values of that elements?

Asked By: user2988577

||

Source

Answer 1

There are three steps in hierarchical agglomerative clustering (HAC):

Quantify Data (metric argument)
Cluster Data (method argument)
Choose the number of clusters

Doing

z = linkage(a)

will accomplish the first two steps. Since you did not specify any parameters it uses the standard values

metric = 'euclidean'
method = 'single'

So z = linkage(a) will give you a single linked hierachical agglomerative clustering of a. This clustering is kind of a hierarchy of solutions. From this hierarchy you get some information about the structure of your data. What you might do now is:

Check which metric is appropriate, e. g. cityblock or chebychev will quantify your data differently (cityblock, euclidean and chebychev correspond to L1, L2, and L_inf norm)
Check the different properties / behaviours of the methdos (e. g. single, complete and average)
Check how to determine the number of clusters, e. g. by reading the wiki about it
Compute indices on the found solutions (clusterings) such as the silhouette coefficient (with this coefficient you get a feedback on the quality of how good a point/observation fits to the cluster it is assigned to by the clustering). Different indices use different criteria to qualify a clustering.

Here is something to start with

import numpy as np
import scipy.cluster.hierarchy as hac
import matplotlib.pyplot as plt


a = np.array([[0.1,   2.5],
              [1.5,   .4 ],
              [0.3,   1  ],
              [1  ,   .8 ],
              [0.5,   0  ],
              [0  ,   0.5],
              [0.5,   0.5],
              [2.7,   2  ],
              [2.2,   3.1],
              [3  ,   2  ],
              [3.2,   1.3]])

fig, axes23 = plt.subplots(2, 3)

for method, axes in zip(['single', 'complete'], axes23):
    z = hac.linkage(a, method=method)

    # Plotting
    axes[0].plot(range(1, len(z)+1), z[::-1, 2])
    knee = np.diff(z[::-1, 2], 2)
    axes[0].plot(range(2, len(z)), knee)

    num_clust1 = knee.argmax() + 2
    knee[knee.argmax()] = 0
    num_clust2 = knee.argmax() + 2

    axes[0].text(num_clust1, z[::-1, 2][num_clust1-1], 'possiblen<- knee point')

    part1 = hac.fcluster(z, num_clust1, 'maxclust')
    part2 = hac.fcluster(z, num_clust2, 'maxclust')

    clr = ['#2200CC' ,'#D9007E' ,'#FF6600' ,'#FFCC00' ,'#ACE600' ,'#0099CC' ,
    '#8900CC' ,'#FF0000' ,'#FF9900' ,'#FFFF00' ,'#00CC01' ,'#0055CC']

    for part, ax in zip([part1, part2], axes[1:]):
        for cluster in set(part):
            ax.scatter(a[part == cluster, 0], a[part == cluster, 1], 
                       color=clr[cluster])

    m = 'n(method: {})'.format(method)
    plt.setp(axes[0], title='Screeplot{}'.format(m), xlabel='partition',
             ylabel='{}ncluster distance'.format(m))
    plt.setp(axes[1], title='{} Clusters'.format(num_clust1))
    plt.setp(axes[2], title='{} Clusters'.format(num_clust2))

plt.tight_layout()
plt.show()

Gives
enter image description here

Answered By: embert

Tutorial for scipy.cluster.hierarchy

Question:

Answers: