Unsupervised Learning

So far, we've had a cheat sheet. We knew which dots were "Dogs" and which were "Wolves". This is called Supervised Learning.

But what if you just have a bucket of Legos? No labels. No instructions. Just chaotic plastic.

You start grouping them yourself:

• "These are red."
• "These are blue."
• "These are weird tiny heads."

This is Clustering. And the most famous algorithm is K-Means.

How K-Means Works

Imagine a high school cafeteria.

1. Selection: You pick K random tables (centroids). Let's say K=3 (Jocks, Nerds, Goths).
2. Assignment: Every student sits at the table they are most similar to.
3. Update: The "center" of the table shifts to be indeed in the middle of everyone sitting there.
4. Repeat: People move tables if the center shifts away from them.
5. Eventually, everyone settles down.

Interactive Demo: The Party Mixer

You have to tell the algorithm K (how many groups you want).

• K=2: Two big blobs.
• K=10: Ten tiny cliques.

The Code (Python)

from sklearn.cluster import KMeans
 
# 1. The Sorter
# n_clusters=3 means we want 3 groups
kmeans = KMeans(n_clusters=3)
 
# 2. Fit (Notice we only give X, no y!)
# The model doesn't know the answers. It finds them.
kmeans.fit(X_data)
 
# 3. Get the labels
groups = kmeans.labels_
# Output: [0, 1, 0, 2, 1, ...]

The Problem

How do you choose K? If K is too small, everything is mashed together. If K is too big, every person is their own group. There is a method called the "Elbow Method" to find the best K, but honestly, often you just guess and see what looks pretty.

Summary

K-Means is like organizing a messy room by throwing similar things into piles. It doesn't know what the pile is (it won't label it "Socks"), but it knows they belong together.

Next up: The engine that powers almost all modern ML. The Blind Hiker.