Unsupervised Learning: Hands-on with K-means Clustering

Before moving on to the practical application, let's refresh our memory with a recap of K-means clustering. K-means clustering is an integral method in unsupervised learning. The main principle of K-means clustering is quite simple: it groups data points into distinct clusters based on their mutual distances to minimize the variance, also known as inertia, within each cluster.

We will now apply K-means clustering to a well-known dataset: the Iris dataset.

The Iris dataset, as we've discussed in previous lessons, consists of measurements taken from 150 iris flowers across three distinct species. Imagine being a botanist searching for a systematic way to categorize new iris flowers based on these features. Doing so manually would be burdensome; hence, resorting to machine learning, specifically K-means clustering, becomes a logical choice!

Let's load this dataset using the sklearn library in Python and convert it into a pandas DataFrame:

from sklearn.datasets import load_iris
import pandas as pd

iris = load_iris()
iris_df = pd.DataFrame(iris.data, columns=iris.feature_names)
iris_df.head()

We're now going to implement K-means clustering on the Iris dataset. For this, we'll use the KMeans class from sklearn's cluster module. To keep our initial implementation straightforward, let's focus on just two dataset features: sepal length and sepal width.

from sklearn.cluster import KMeans

# Assigning the features for our model
X = iris_df.iloc[:, [0, 1]].values

# Defining the KMeans clustering model
kmeans = KMeans(n_clusters=3, init='k-means++', max_iter=300, n_init=10, random_state=0)
y_kmeans = kmeans.fit_predict(X)

In the above block:

• n_clusters: stipulates the number of clusters to form.
• init: sets the method for initialization. The "k-means++" method selects initial cluster centers intelligently to hasten convergence.
• max_iter: limits the maximum number of iterations for a single run.
• n_init: specifies the number of times the algorithm runs with different centroid seeds.
• We have left out parameters like tol and algorithm:
  • tol is the tolerance with regard to inertia required to declare convergence. We did not include this to use the default tolerance.
  • algorithm specifies the algorithm to use. The classical EM-style algorithm is "full", the Elkan variant is more efficient but is not available for sparse data. To keep things simple, we chose not to specify this option.