Original subtitles

Instructor: It wouldn't be data science

if there wasn't this very important topic,

problems with, issues with, or limitations of X.

Well, let's look at the pros and cons of K-means clustering.

The pros are already known to you

even if you don't realize it.

It is simple to understand and fast to cluster.

Moreover, there are many packages that offer it,

so implementation is effortless.

Finally, clustering it always yields a result.

No matter the data, it will always spit out a solution

which is great.

Time for the cons.

We will dig a bit into them

as they are very interesting to explore.

Moreover, this lecture will solidify your understanding

like no other.

The first con is that we need to pick K.

As we already saw, the elbow method fixes that,

but it is not extremely scientific per se.

Second, K-means is sensitive to initialization.

That's a very interesting problem.

Say that these are our points.

If we randomly choose the centroids here and here,

the obvious solution is one top cluster

and one bottom cluster.

However, clustering the points on the left in one cluster

and those on the right in another

is a more appropriate solution.

Now imagine the same situation,

but with much more widely spread points.

Guess what?

Given the same initial seeds,

we get the same clusters because that's how K-means works.

It takes the closest points to the seeds.

So if your initial seeds are problematic,

the whole solution is meaningless.

The remedy is simple. It is called K-means++.

The idea is that a preliminary iterative algorithm is ran

prior to K-means to determine the most appropriate seeds

for the clustering itself.

If we go back to our code,

we will see that sklearn employs K-means++ by default,

so we are safe here,

but if you are using a different package,

remember that initialization matters.

A third major problem

is that K-means is sensitive to outliers.

What does this mean?

Well, if there is a single point

that is too far away from the rest,

it will always be placed in its own one-point cluster.

Have we already experienced that?

Well, of course we have.

Australia was the sole cluster in almost all the solutions

we had for our country clusters example.

It is so far away from the rest of the countries

that it is destined to be in its own cluster.

The remedy, just get rid of outliers prior to clustering.

Alternatively, if you do the clustering

and spot one-point clusters, remove them and cluster again.

A fourth con, K-means produces spherical solutions.

This means that on a 2D plane that we have seen,

we would more often see clusters that look like circles

rather than elliptic shapes.

The reason for that is that we are using Euclidean distance

from the centroid.

This is also why outliers are such a big issue for K-means.

Finally, we have standardization.

Oh, good old standardization.

Let's leave that for the next lesson, shall we?

Thanks for watching.