All language subtitles for 07 - Feature Selection, Feature Learning, and Feature Extraction.en

[Autogenerated] Let's dive into a little

more detail on each of these components or

future engineering, starting with feature

selection. Feature selection involves

choosing the best subset from within. An

existing set of features are ex variables

without substantially transforming or

changing the features in any manner when

you're building and training on machine

learning model. When would you use feature

selection? Let's say you have many ex

variables present in your data, and not

all of these X variables contained

information. During exploratory data

analysis, you found that most off your

features are ex variables contained little

information. They're not relieved relevant

to your problem. But there are a few

features. Are ex variables that are very

meaningful and have high predictive power,

and you find that these meaningful

variables are independent of each other.

You will then use feta selection toe only.

Choose those meaningful variables to train

your model. Feature. Selection techniques

can be divided into three broad

categories. The first of these are filter

method, where you apply a statistical

technique to extract the most relevant

features. You can use embedded methods

where you build a machine learning Marty

that assigns importance to the different

features and select those features that

are the most important. And the third

technique that you can apply is rapper

methods rapper methods like somewhere

between filter method and embedded

methods. Here you train a number of

different candidate models on a subset of

features and choose that subset of

features that produces the best model.

Filter methods for future selection are

very you applied statistical techniques to

select those features that are the most

relevant feature. Selection is completely

independent off. Actually building and

training the model with fatal metals,

you'll use techniques such as sky Square.

Analysis on over analysis or mutual

information to determine board features

are relevant to the target that you're

trying to predict. Embedded method for

feature selection involved training a

machine learning model. Now not all models

assigned importance measures to features

Relevant features can be selected by

training a machine learning algorithm,

which under the hold will sort features by

important and assigning importance or

relevant score. Tow each feature. Examples

of such models are last regression on

decision Trees feature selection using

embedded metadata embedded within the

mortal training fees. That's why they're

called embedded methods when you perform

feature selection using rapper method you

build a number of different candidate

models on each of these candidate models

are built on different subsets off your

features. You then choose that fetus

upset, which gives you the best model.

Let's go back to the different components

of feature engineering that people

discussing earlier and let's move on to

discussing feature learning. Feature

learning is where you rely on machine

learning algorithms rather than human

experts to learn the best representation

off. Complex leader. This is especially

useful when you're working with data such

as images or videos. Feature Learning is

also often referred to as representation

learning. Now you can have fetal learning

techniques that are supervising nature

when you're working with a label. Corpus

off data. Neural networks are classic

example. Off, supervise feature learning.

Then you feed data into neural networks.

You typically don't highlight significant

features. You just feed all off the data

in and neural networks find out what late

and features are important or significant

Supervise. Future learning is an extremely

important technique because it greatly

reduces the need for expert judgment. When

the selection of your features rely on

humans, that technique will simply not

skill. It's much better to have an

automated solution. Such a supervised

feature learning and the fact that new

neural networks can learn significant

features in your data is an important

difference between neural network and deep

learning techniques and traditional

machine learning based systems.

Traditional ML systems rely on experts to

decide what features to pay attention to.

On the other hand of representation, ML

based systems figure out by themselves.

What features are important or

significant. Neural networks are examples

off representation MLB systems. One thing

you'll see when you work in the real ball

that it's really hard to get a label

corpus. Most of the data out in the real

world tends to be unlabeled, and

thankfully, there are future learning

techniques that can work with an unlabeled

corpus as well. These techniques are for

tow US unsupervised feature learning on

surprise, indicating the absence of label

data. In order to learn patterns and

unlabeled data, you can apply clustering

techniques. Clustering will allow you to

find a logical groupings that exist in

your data. If you're working with image

data and unsupervised technique for

feature learning is dictionary Learning

Dictionary Learning learns sparse

representation. Zoff dense features such

as with images. If you're working with

deep learning, specifically neural

networks, you can use auto and quarters to

extract latent significant representations

off your data. Now you might have machine

learning models that can't work directly

with raw features, which worked better

with derived features. And that's their

feature. Extraction comes in feature

extraction differs from feature selection

that we discussed earlier in that the

input features are fundamentally

transformed into derived features. The

delight features are often unrecognisable

and may be hard to interpret. These

derived features might represent patterns

that exist in your data, which are not

intuitively understandable. Feature

extraction techniques exists for data off

all kinds of your working with images. You

can use key points and descriptors to

represent interesting areas in your image.

If you're working with simple numeric

data, feature extraction might involve

reorienting your data to be protected

along new axes, such as what we do when we

find the principle components for me to

seize or don't feel working with text data

and natural language processing, you might

choose to represent the worlds within your

document using their DF idea of scores.

The stands for Tom Frequency inverse

document frequency and this is a score

that represents how significant a

particular word is in a particular

document and across the entire corpus. Now

it so happens that when you perform

feature extraction, this also leads to

dimensionality reduction. It reduces the

number of dimensions which are required to

express their data. So many off. The

feature extraction techniques that we

discussed earlier also happened to be

dimensionality reduction techniques.

However, when you use these techniques for

feature extraction, the explicit objective

is to the express your feature in a better

form, not to reduce the number of ex columns or features.