Clusters vs Known Labels

Part of the Unsupervised Learning section of Coddy's Machine Learning journey. Lesson 13 of 51.

When labels exist but k-means never saw them, check what it found: pd.crosstab counts each pair of cluster and label.

labels = KMeans(n_clusters=3, n_init=10, random_state=0).fit_predict(X)
print(pd.crosstab(labels, kind))

For 150 coffee beans of 3 kinds, clustered on length and caffeine: one row per cluster, one column per kind:

A 3 by 3 table of counts, clusters down and coffee kinds across: arabica, liberica, robusta. Cluster 0 holds 49 robusta beans; cluster 1 holds 48 arabica, 3 liberica and 1 robusta; cluster 2 holds 47 liberica and 2 arabica

Each cluster is mostly one kind; only 6 beans strayed. And k-means never saw a label.

Clusters are not classes, though. A close match means the classes differ in the data. A poor match is a finding, not a bug.

Challenge

Medium

Do k-means clusters find the iris species? The starter scales the flowers, keeps their names in species, reads k and clusters them.

Crosstab the clusters against the species.
For each species, print its largest count in one cluster, out of its total.
Count the clusters that hold a single species.

For another garden and k = 3:

setosa: 48 of 50 share a cluster
versicolor: 41 of 50 share a cluster
virginica: 33 of 50 share a cluster
Clusters with one species: 0 of 3

Try it yourself

import pandas as pd
from sklearn.datasets import load_iris
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans

# 150 iris flowers, scaled, and the species name of each, hidden from k-means
iris = load_iris()
X = StandardScaler().fit_transform(iris.data)
species = iris.target_names[iris.target]

k = int(input())
labels = KMeans(n_clusters=k, n_init=10, random_state=0).fit_predict(X)

# Write your code below