Original subtitles

Instructor: The most common problem

when working with numerical data

is about the difference in magnitudes

as we mentioned in the first lesson.

An easy fix for this issue is standardization.

Other names by which you may have heard this term

are feature scaling and normalization.

However, normalization could refer

to a few additional concepts even within machine learning

which is why we'll stick with the term standardization

and feature scaling.

Standardization or feature scaling

is the process of transforming the data

we are working with into a standard scale.

A very common way to approach this problem

is by subtracting the mean

and dividing by the standard deviation.

In this way,

regardless of the data set,

we will always obtain a distribution with a mean of zero

and a standard deviation of one,

which could easily be proven.

Let's show that with an FX example.

Say our algorithm has two input variables,

Euro dollar exchange rate and the daily trading volume.

We have three days worth of observations.

First day, 1.3 and 110,000,

second day, 1.34 and 98,700,

and the third day, 1.25 and 135,000.

The first value shows the Euro dollar exchange rate,

while the second one shows the daily trading volume.

Let's standardize these figures.

We standardize the Euro dollar exchange rates

regarding the other Euro dollar exchange rates.

So, we look at 1.3, 1.34 and 1.25.

The mean is 1.3,

while the standard deviation 0.045.

Going through the above mentioned transformation,

these values become 0.07, 0.96 and -1.03 respectively.

Standardizing trading volumes,

we obtain -0.25, -0.85 and 1.1.

In this way,

we have focused figures of very different scales

to appear similar.

That's why another name for standardization

is feature scaling.

This will ensure our linear combinations

treat the two variables equally.

Also, it is much easier to make sense of the data.

The transformation of trading volumes

allowed us to transform the volumes

from 110,000, 98,700 and 135,000 to -0.25, -0.85 and 1.1.

In this way,

the third term is considerably higher than the average,

while the first one is around the average.

We can confidently say that 135,000 trades per day

is a high figure,

while 98,700 is low.

Please disregard the simplification

of having just three observations.

That's just an example.

Besides standardization,

there are other popular methods, too.

We will shortly introduce them

without going too much in detail.

Initially, we said that normalization

refers to several concepts.

One of them, which comes up in machine learning

often consists of converting each sample

into a unit length vector using the L1 or L2 norm.

Another pre-processing method is PCA

standing for principal components analysis.

It is a dimension reduction technique

often used when working with several variables

referring to the same bigger concept or latent variable.

For instance,

if we have data about one's religion,

voting history,

participation in different associations,

an upbringing,

we can combine these four

to reflect his or her attitude towards immigration.

This new variable will normally be standardized

in a range with the mean of zero

and a standard deviation of one.

Whitening is another technique frequently used

for pre-processing.

It is often performed after PCA

and removes most of the underlying correlations

between data points.

Whitening can be useful when conceptually,

the data should be uncorrelated.

But that's not reflected in the observations.

We can't cover all the strategies

as each strategy is problem specific.

However, standardization is the most common one

and is the one we will employ

in the practical examples we will face in this course.

In the next lesson,

we will see how to deal with categorical data.

Thanks for watching.