Nearest Neighbours
You are the teacher! Put apples, lemons and pears on the table. Then drag the mystery fruit around: the robot guesses it by asking the fruit that sit closest to it.
I can't taste or smell. I only know two things about a fruit: how round and how yellow it is. So every fruit is a spot on my table. Drag the big ? around and I'll guess what lives there!
Tap the table to drop the fruit you picked · drag any fruit to move it · drag the big ? to ask the robot
Snapshots while learning
A photo of my map after 1, 2, 4, 8… examples. More examples = a smarter map. Tap one to look back.
How it works
Examples with answers
Every fruit you drop comes with a name tag. The name tag is the "answer".
Turn fruit into spots
Roundness goes left-right, yellowness goes up-down. Every fruit becomes a spot on the table.
Ask the neighbours
For a new fruit, find the K fruit sitting closest to it.
Most votes wins
Each neighbour votes for its own name. The name with the most votes is the guess!
What just happened?
- You gave the robot examples, each with the right answer.
- The robot didn't do any hard thinking — it just remembered every example.
- To guess a new fruit, it asked the closest examples and counted their votes.
- More examples made the map smarter and the test score went up.
- Asking 1 neighbour is jumpy; asking many is smoother but can miss small groups.
For grown-ups: what's really going on
This is k-Nearest Neighbours (k-NN) classification with two features (roundness, yellowness) and three labels.
- k-NN is a lazy / instance-based learner: "training" is just storing examples. Prediction finds the
kclosest stored points by Euclidean distance√((x₁−x₂)² + (y₁−y₂)²)and takes a majority vote. - The tinted background is the decision boundary; colour strength shows how unanimous the vote is.
ktrades variance for bias:k = 1memorises noise (jagged islands), largeksmooths everything and can out-vote small classes. Oddkavoids ties with two classes.- The test score is accuracy on a held-out test set drawn from the same distribution; the chart is a learning curve (accuracy vs. training-set size).
- In real use, features must be on comparable scales (normalise them), and k-NN gets slow and weak in high dimensions (the curse of dimensionality).
Further reading: An Introduction to Statistical Learning, §2.2.3 · scikit-learn "Nearest Neighbors" guide.