linkedin.

com/in/vikrantkumar95

K-Means Clustering
Clearly Explained
 What is Unsupervised Learning?
Before we jump into what K-Means Clustering is, let see what the
difference between Unsupervised and Supervised Learning is:

Supervised Learning

Trained Model Supervised Learning usually consists of an

algorithm that learns from labeled data. The
model learns to recognize characteristics
based on the training data and the label (Eg.
Predicting Property prices based on size,
location etc). It then is able to predict the label
of unlabelled data on based on those
characterics.

Unsupervised Learning
Unsupervised Learning is a type of machine
learning where an algorithm learns patterns,
structures, or relationships in data without any
labeled outputs. The goal is to uncover hidden
insights or organize data into meaningful
groups based on similarities. K-Means
Clustering is a type of Unsupervised Learning
Alogrithm.

linkedin.com/in/vikrantkumar95
 What is K-Means Clustering
As we mentioned in the previous page, K-Means Clustering is a
Unsupervised Learning method. It works very simply:

Divides data into K distinct clusters

Groups similar data points based on their features
Ensures data points in a cluster are closer to each other than to
points in other clusters

Unclustered Data First Cluster, could be better?

Looks Much Better

Since it’s an unsupervised method, we don’t have any labels to train a

model on a training set. The K-means algorithm iterates by creating
clusters (we give the number of clusters) and then measuring if it’s
good. So how does it create clusters and how does it measure the
“goodness” of a cluster? Let’s see how it works!

linkedin.com/in/vikrantkumar95
 How does it work?
Let’s start with a one-dimensional data set : [1, 3, 4, 8, 10, 12]

1 3 4 8 10 12

Step 1: Decide on the number of clusters (K)

We’ll create 2 clusters here. So we take K = 2.

Step 2: Decide on the number of clusters (K)

Randomly select two initial “Centroids” (equal to k) from the data points.
Here centroid simply means the mean of all the points in a cluster. Since
we don’t have a cluster yet, we will select 2 data points. The initial
Centroids we picked here are : [3, 10]

Step 3: Calculate the Distance of Each Centroid from the

data points

Data Distance to Distance to Assigned

Point Centroid 3 Centroid 10 Cluster

1 2 9 3

3 0 7 3

4 1 6 3

8 5 2 10

10 7 0 10

12 9 2 10

linkedin.com/in/vikrantkumar95
 How does it work?
We calculated the distance in the table previously. Since it’s one
dimensional data, the distance is just the difference between the
centroid and the data point. We assigned each point to the cluster
with closest centroid

1 3 4 8 10 12

Step 4: Update the Centroids

Now that we have our two clusters, we calculate new Centroids. It’s
simply the mean of all the points in the cluster.

1+3+4
New Centroid for Cluster 1 = = 2.67
3

8 + 10 + 12
New Centroid for Cluster 2 = = 10
3

These two points are now our new Centroids [2.67, 10]
We then repeat steps 3 and 4 again. We will c

1 2.67 3 4 8 10 12

linkedin.com/in/vikrantkumar95
 How does it work?
Step 5: Repeat Steps 3 - 4

Data Distance to Distance to Assigned

Point Centroid 2.67 Centroid 10 Cluster

1 1.67 9 2.67

3 0.33 7 2.67

4 1.33 6 2.67

8 5.33 2 10

10 7.33 0 10

12 9.33 2 10

Now from what we can see above:

Even though the Centroids changed (atleast one did), the clusters
remained the same [1,3,4] & [8,10,12].
If we were to repeat steps 3 and 4 again, we would get the same
result since the we’d get the same Centroids again. The alorithm
thus stops here and we have the best presumed cluster.
The alogrith stops when the Centroids don’t change or when it
reaches the max number of iterations.

Now let’s see how it would perform on a 2 Dimensional Dataset!

linkedin.com/in/vikrantkumar95
 How does it work?

Step 1: Let’s create 3 clusters this time. So K = 3

Step 2: With a 2-D data set, we will again start with picking the
Centroids. These will be 3 points randomly selected.

The Red, Green and Yellow points are our initial centroids.

linkedin.com/in/vikrantkumar95
 How does it work?
Step 3: Similar to when we had 1-D data, we have to calculate
the distance of each point from the centroids and then put them
in the cluster with the closest centroid. Now, to calculate the
distance in 2-D and more dimensions, we use something called
Euclidean Distance d

Step 4: After calculating the distances, we would then assign

each point to the cluster belonging to the Centroid that is at the
least distance from it.

These are our new clusters. We now update the centroids again,
by calculating the mean of all the points in a particular cluster.

x1 + x2 + x3 +x4 y1 + y2 + y3 +y4
New Centroid for Cluster 1 =
,
4 4

We do the same for Cluster 2 and Cluster 3.

linkedin.com/in/vikrantkumar95
 How does it work?
Step 5: Repeat Step 3 & 4

C2
C1

C3

With our new Centroids C1, C2 and C3 we again calculate the

distance of each point from each Centroid and bucket them in
the cluster belonging to the nearest Centroid.

C2
C1

C3

This looks like a much better set of clusters! The algorithm will
continue to create new centroids and changing clusters till either
the centroids don’t change significantly or the max iterations is
reached.

linkedin.com/in/vikrantkumar95
 How do we decide K
A question people often have is, how do we decide on the value
of K? In our examples it was somewhat clear by looking at the
data, how many clusters would be ideal. But if it’s a higher
dimensional data, we will not be able to visualise it. How do we
decide how many clusters should we create?

Since with each iteration of the centroids, the distance between

the points and centroid decreases, we are essentially
decreasing the varations within a cluster. Variation would simply
be the sum of the square distances between the centroid and
the data points.

With each iteration, this varation will decrease. However, at some

point, the decrease in variance will start to get smaller and
smaller.
Decrease in Variation

Elbow Point

1 2 3 4 5
k

We usually choose a K that’s at the “Elbow Point”, above which

there’s diminishing returns in creating more clusters.

linkedin.com/in/vikrantkumar95
 Let’s Summarise
Let’s summarise what K-means Clustering is:

1. What It Is
A simple and popular unsupervised learning algorithm.
Groups data into K clusters based on similarity.

2. How It Works
Step 1: Choose K
Step 2: Initialize K centroids randomly.
Step 3: Assign data points to the nearest centroid.
Step 4: Update centroids as the mean of assigned points.
Step 5: Repeat until centroids stabilize.
3. Goal
Minimize within-cluster variance (make clusters compact).

4. Strengths
Easy to understand and implement.
Works well for structured, spherical data.

5. Challenges
Requires predefining K (number of clusters).
Sensitive to initialization and outliers.
Assumes clusters are spherical and similar in size.

6. Applications
Market Segmentation: Group customers by behavior.
Anomaly Detection: Identify unusual patterns.
Recommender Systems: Cluster similar users or items.
Image Compression: Reduce pixel redundancy.

linkedin.com/in/vikrantkumar95
 Enjoyed
reading?

Follow for
everything Data
and AI!

linkedin.com/in/vikrantkumar95