It's time to revisit the

exciting topic of machine learning.

Previously in this course,

your data science team had to build

a recommendation model using Spark ML,

that you then ran on Cloud Dataproc for scale.

Later in this module, you're

going to be building custom models

yourself with just SQL using BigQuery ML.

But before we jump into the code,

we need to expand our Machine Learning Foundation

and then cover the models and

the key terminology that you need to know

before we set you off to make predictions.

When people think of AI in machine learning,

they generally think of the advanced models

like you saw earlier on Google photos,

in video stabilization, and the

smart reply feature in Gmail.

Yes, later on this course you will build

the image models and unstructured data sets.

But did you know that at Google,

the majority of the models

deployed or models that operate unstructured data.

These aren't your 50 plus layer deep neural networks

that play StarCraft or chess.

They're built on rows and columns of data,

just like you've see experimented

with inside of BigQuery.

So if you have a structured data set

that you think is a good use case for machine learning,

the next step is to find

a model type that is appropriate for your use case.

Add up all the models out there.

What's a good place to start

for you to start prototyping?

Here's a decision tree no pun intended to help guide us.

We'll walk through each of the different branches.

The first question is,

what kind of activity that you're engaging in?

Is there a right answer or a ground truth that

exists in your historical data

that you want to model and predict?

If so, you want to start with supervised learning.

Alternatively, if you're interested in just ruminating

and exploring the data for unknown relationships,

you want to try unsupervised learning

with maybe a clustering model to start.

Unsupervised learning is outside

the scope of this course,

but link you to a few resources to show how you can do it

quickly with a clustering model inside of BigQuery ML.

The majority of the problems we're going to tackle

here are nice three areas.

First, forecasting.

That's like predicting the next month's sales figures

that demand for your product.

Second, classifying.

Like high medium or low

risk events or buy or no buy decisions.

Third, maybe you recommending

something like a product for a given user.

An easy way to tell if you're forecasting or classifying,

is to look at the type of label or special column,

will cover that more later,

of data that you're predicting.

Generally, if it's a numeric datatype like units

sold or profits earned, you're doing forecasting.

If it's a string value,

you're typically doing classification.

This row is either in this class or this other class,

and if you have more than two classes

or buckets like high,

low, medium, you're doing

what's called multi-class classification.

Now, once you have your problem outlined,

it's time to go shopping for models

which will be the tools to help you achieve your goal.

Now there are many different model types

for you to choose from for these problems.

We're recommending you start with

a simpler ones which can still be

highly accurate to see if they meet your benchmark.

By the way, you're ML benchmark is

the performance threshold that you're willing to

accept from your model before you even

allow it to be near your production data.

It's critical that you set your benchmark

before you train your model.

So you can really be truly

objective in your decision-making

to use the model or not.

Now, on to types of models.

For forecasting, try a linear regression.

For classification, try logistic regression.

By the way is called binary logistic regression.

If you have a GIS two classes

or buckets that an observation

could fall into or multi-class if it's more than two.

For recommendations, try matrix factorization which is

a commonly used algorithm for problems involving

a matrix of users and items.

Like your housing model's example,

and here's the complete picture again.

You'll see later with BigQuery ML

that you can just specify

a model type equal to linear regression for example,

and BigQuery handles the rest for you.

What didn't you see here that you might

have heard of in terms of a model type?

There's many different types of models out

there that you may not see on this chart.

More complex models like deep neural networks,

decision trees,

random forests are also available for modeling.

You'll even build a custom model

using neural architecture search to build

a deep neural network later on in this course and you'll

do so without using any code that's what Auto ML.

It's my overall recommendation that even if you

know how to build advanced models,

that you start with the simpler ones first.

Because they often trained

faster and they give you an indication

of whether not ML is

even a viable solution for your problem.