Original subtitles

Instructor: Hey, welcome back

to this market segmentation problem.

We were just investigating an elbow.

There isn't a clear tip of the elbow.

I can see three or four tips which are worth trying,

at two, three, four, and five clusters.

Here's the limitation of the elbow method.

We can see the change in WCSS with the increase

in the number of clusters,

but we don't really know which solution is the best one.

All right, let's try two clusters.

We already discussed qualitatively

that this is probably a suboptimal solution,

but it is worth inspecting the difference

with standardized variables.

We'll declare a variable called kmeans_new

equal to KMeans of two.

Next, we will fit the x_scaled data.

Finally, we will create a new data frame

called clusters_new, containing the values from x.

Then the column cluster_pred

will contain the predicted clusters

from this new clustering solution with the scaled x.

Here is a crucial moment.

The data frame contains the original values,

but the predicted clusters are based

on the solution using the standardized data.

This is very important.

We will plot the data without standardizing it,

but the solution itself is the standardized one.

Let me show you by plotting the data.

By keeping the original x-axis, we get the intuition,

how satisfied were the customers.

If we plot the standardized values, we would be deceived.

The middle parts of the two graphs are different.

This one is 5.5, and on the standardized graph,

the midpoint zero actually corresponds

to the mean of the variable, or 6.4.

Let that sink in for a second.

If you wish, you can rewind to be sure you got that right.

We will now continue with the solution.

We can see two clusters,

as we specified the number to be two.

No surprise here.

What's different, though, is the clusters themselves.

Comparing this result with the previous one,

we can clearly see

that both dimensions were taken into account.

Moreover, these two clusters coincide

with our initial speculations,

that those two would be the result of k equals two.

Okay, great, we are now much more confident

that standardization is generally a good thing.

However, the problem is not solved yet.

This two-cluster solution does not make a whole lot

of sense, as we discussed before, but it's a good start.

Let's name the two clusters.

One contains people with low loyalty and low satisfaction,

so we can call these people alienated.

By the way, naming your clusters is very important.

In unsupervised learning, clustering included,

the algorithm will do the magic,

but then we step in to interpret the result.

My feeling here is to call them the alienated cluster,

as they are dissatisfied and not loyal.

No wonder, it's unlikely they'll be back to our shop.

As for the other cluster, it is so heterogeneous

that I'd call it the everything else cluster.

All right, let's get back to the elbow.

Noteworthy tips of the elbow are also three, four, and five.

I'll try them one after the other.

With our well-parameterized code,

we can just change the number of clusters in the first line,

and rerunning the code would do the trick.

Let's try with three clusters.

That's the result.

We have the alienated cluster once more.

That's a good sign.

It shows us that we were right

in concluding that it is a cluster of its own,

while the everything else cluster is now split into two.

I'd call this group the supporters.

They are not particularly happy

with the shopping experience, but they like the brand

and wanna keep coming back.

Note that there are not that many of them.

It is a small cluster.

Finally, the third cluster is called, well,

the all that's left cluster, I guess.

We can't really name it as it is still very much mixed.

What happens next?

Let's check out a four-cluster solution.

We have the alienated and the supporters clusters,

and now these two new ones can also be named, finally.

The upper right one consists of clients

that are satisfied and loyal.

These are our fans, the core customers.

Eventually, we hope that all the points

on this graph turn into fans,

but we will elaborate on this later.

Let's name the last cluster.

We have people who are predominantly satisfied

but not loyal, and some of them are actually disloyal.

A term I've seen somewhere

to describe such customers is roamers.

They like your brand, but they are not very loyal to it.

We have all been there for some brand.

Okay, this solution is definitely the best one

we've seen so far.

Here's where it stood on the elbow graph,

but how about we try with five clusters?

The alienated, the supporters,

and the fans remain unchanged.

These people here look like the roamers from before.

Finally, these clients are almost in the middle

of our standardized graph.

They almost neutral on the loyalty feature

but are generally satisfied.

They are also roamers.

This solution actually split the roamers

into two subclusters, those that are extremely satisfied

and those that are just satisfied,

so there isn't much value added to our segmentation.

We can carry on with as many clusters as we want,

but from now on,

we would just further segment the four core clusters.

Let's finish off with nine clusters.

Similar to what we had a second ago,

many of the clusters were further segmented.

It is extremely hard to name all of them,

and even if we do,

we will probably need to use a lot of adjectives.

For instance, the alienated cluster is split into two,

the very alienated cluster

and the moderately alienated cluster.

As you can imagine, there is not much to gain

by using such a fragmented.

In my mind, the four and five-cluster solutions

were the best ones.

Which one you want to use depends on the problem at hand.

Okay, in the next lesson, we will see what we can do

with this new information.

Thanks for watching.