Week 1 - Introduction and Classification
Intro to analytics :
Hi. Welcome to introduction to analytics modeling. Georgia Tech's course ISYE 6501.
I'm Joel Sokol, Director of the Master of Science in Analytics degree at Georgia Tech and a
professor in Georgia Tech's Stewart School of ISYE.
In almost all of that work, a critical key to success has been analytics modeling. Selecting
and specifying the right analytics model or models to combine into a solution, compiling,
building and or forecasting the necessary data sets, and then interpreting the model's
output to make suggestions that match the organizations needs, priorities, and structure.
And that's what this course is about. As the name says, it's an introduction to analytics
modeling. Analytics can help answer lots of important types of questions.
It can answer descriptive questions.
Questions that ask for explanations of what happened . Like how much effect did a new
system for scanning carry-on luggage have on airport security wait times? When did the
reliability of
a critical component of a piece of manufacturing equipment dropped below an acceptable
level?
Which sets of customers are most alike in their buying patterns? And what factors are most
important in determining customer similarity? Is a new medical treatment better than what's
currently being used?
Analytics can also answer predictive questions, questions that ask what's going to happen in
the future.
How much worldwide demand will there be for crude oil next year, or five or 10 years from
now? What will Google stock price be a year from now? And how much uncertainty is there in
that estimate? How likely is it that a person with certain characteristics will eat in my restaurant
chain if I sent him a coupon or vote for the political candidate I support, or click on a specific
banner ad on a website.
1
,And analytics can also answer prescriptive questions. Questions that ask what action or
actions would be best? How should traffic lights be timed and synchronized to minimize
delays downtown when there's a major sporting event.
When and where should giant oil tankers make pickups and deliveries all around the world?
What strategies can an airline needs to get passengers to their destinations quickly before,
during and after a big snow storm.
Finally, sometimes analytics is useful for answering even more general questions like we just
installed technology to track all detailed information that we never had before, so how do we
monetize it?
How can we use that data to create value for our company? We'll talk about all of these types
of questions as we go through the course. In both discussions of when analytics can help you
answer them, and in lessons on how to use analytics models and methods to help find the
answers you're looking for. In addition to viewing lessons from me, you'll have a chance to
contribute your own thoughts and ideas too, as part of the discussion forums, homework
assignments, and the peer grading process.
So that's the analytics part of this introduction to analytics modeling course, now let me say a
little bit about modeling. When most people think of modeling, this is the thing they think of.
Luckily for all of us, that's not what I'm going to be doing on these videos. You wouldn't want
to see it and I wouldn't want to do it.
In analytics, modeling means taking a real life situation and expressing the key parts of that
situation in terms of math. So we can analyze the math and then turn the mathematical
analysis back into a real life answer or solution or recommendation.
What is modeling :
In this short lesson, I'm going to try to head off a potential source of confusion. In analytics,
people use the word model to mean three slightly different things.
In the course introduction, we talked about how modeling means taking a real-life situation
and expressing the key parts of that situation in terms of math. So we can analyze the math
and then turn the mathematical analysis back into a real-life answer or solution or
recommendation.
The mathematical expression of the real-life situation is called a model.
2
,And confusingly, there are at least three different ways of using the word model and analytics. If
I say that I'm going to use regression to predict the delivery date of packages accompanying
ships out then regression is referred to as a model.
Anyway, if I plan to use regression to predict the delivery dates, then regression is the model
I'm using. And if I go one level of detail lower and say that I'm going to use regression to
predict
delivery dates based on a package's size, weight, and distance between where it's being
shipped from and where it's being shipped to then all that detail, the use of regression and the
use of size, weight, and distance is also called a model.
And if I go even one more step and use regression to find that the time for a package to be
delivered is approximately equal to 37 plus 81 times the sum of its linear dimensions plus
76 times its weight plus 4 times the delivery distance, then that specific equation is also called
a model.
Intro to Classification :
In this lesson, we'll talk about classification, including what it means, examples of when it's
important, and a simple demonstration of how classification works. Then in later lessons, we'll
see a basic model for solving classification problems.
Classification in analytics has the same meaning as it does in everyday life putting things into
categories.
The simplest examples of classification are when there are two categories, which are often
just yes and no. For example, a bank might want to differentiate between loan applications
who will
fully repay their loan and those who won't. A security agency might want to differentiate
between a regular person and a potential terrorist. An automated email filter might need to
differentiate between real and spam email. A legal document system might be designed to
differentiate between documents that are relevant or irrelevant to a certain case.
In my own research, we're working with the CDC on a project to increase the number of
organs that can be given to people in need of a transplant. And one of the questions we use
analytics to answer is, is this organ safe to transplant or does it carry a deadly infectious
disease?
3
, Of course, one answer would be to subject the organ to a laboratory test, but if the donor was
recently infected, the test might not come out positive yet. So we need the analytical approach.
In each case, we can use classification models to put applicants, people, documents, livers, etc,
into one of the categories. And as you can imagine, having more than two categories as well is
possible. For example, a political consultant might want to differentiate between supportive
voters, opposition voters, and undecided voters or a paleontologist might want to differentiate
between many different species of dinosaurs to determine which one in newfound bone belongs
to. Each of these classification questions will require some data in order to get answers. For
example, for loan applicants, a bank might collect data on income, credit history, age, family
size, assets, liabilities, and more.
Based on those attributes of previous loan recipients and the bank's observation of whether
each loan was repaid or not, the bank can then build a model to help classify future
applicants. Let's see how this would look graphically. Suppose the bank is trying to decide
whether or not to give loans to applicants using the applicants credit scores and incomes. So
we can plot each past recipients information on this graph. The horizontal axis will show
credit score and the vertical axis will show household income.
For example, someone with a credit score of 730 and an income of $76,000 a year would be
this data point. If they repaid the entire loan, we'll mark the data point in blue. If they defaulted,
we'll mark it in red. Of course, in real life, there might be other factors, each of which would be
another dimension, length of time as a customer of this bank might be a third factor, number of
dependents might be a fourth, total assets and liabilities would be a fifth and so on.
But as my wife, the artist, could tell you, I have enough trouble drawing clear two-dimensional
pictures. Three- dimensions is really iffy for me and I have no idea how to draw it on four more.
So we'll use this two-dimensional picture to see how classification works. As you can see, we
can draw a line like this that separates between the blue points above and the red points below.
So if we want to use this data to decide whether to offer a loan to a new applicant, we can just
see where the new applicants data point would be relative to the line.
An applicant up here, would be in the blue zone. So we might want to give them a loan.
An applicant down here, would be in the red zone so we might not want to give them a loan.
But how do we know that we've drawn the right line? There are lots of lines, infinitely many,
in fact, that would also separate the blue points from the red ones.
Choosing a Classifier :
In this lesson, we will see how to think about tradeoffs and classification problems graphically
to build your intuition before we get to the underlying mathematical models.
4
Intro to analytics :
Hi. Welcome to introduction to analytics modeling. Georgia Tech's course ISYE 6501.
I'm Joel Sokol, Director of the Master of Science in Analytics degree at Georgia Tech and a
professor in Georgia Tech's Stewart School of ISYE.
In almost all of that work, a critical key to success has been analytics modeling. Selecting
and specifying the right analytics model or models to combine into a solution, compiling,
building and or forecasting the necessary data sets, and then interpreting the model's
output to make suggestions that match the organizations needs, priorities, and structure.
And that's what this course is about. As the name says, it's an introduction to analytics
modeling. Analytics can help answer lots of important types of questions.
It can answer descriptive questions.
Questions that ask for explanations of what happened . Like how much effect did a new
system for scanning carry-on luggage have on airport security wait times? When did the
reliability of
a critical component of a piece of manufacturing equipment dropped below an acceptable
level?
Which sets of customers are most alike in their buying patterns? And what factors are most
important in determining customer similarity? Is a new medical treatment better than what's
currently being used?
Analytics can also answer predictive questions, questions that ask what's going to happen in
the future.
How much worldwide demand will there be for crude oil next year, or five or 10 years from
now? What will Google stock price be a year from now? And how much uncertainty is there in
that estimate? How likely is it that a person with certain characteristics will eat in my restaurant
chain if I sent him a coupon or vote for the political candidate I support, or click on a specific
banner ad on a website.
1
,And analytics can also answer prescriptive questions. Questions that ask what action or
actions would be best? How should traffic lights be timed and synchronized to minimize
delays downtown when there's a major sporting event.
When and where should giant oil tankers make pickups and deliveries all around the world?
What strategies can an airline needs to get passengers to their destinations quickly before,
during and after a big snow storm.
Finally, sometimes analytics is useful for answering even more general questions like we just
installed technology to track all detailed information that we never had before, so how do we
monetize it?
How can we use that data to create value for our company? We'll talk about all of these types
of questions as we go through the course. In both discussions of when analytics can help you
answer them, and in lessons on how to use analytics models and methods to help find the
answers you're looking for. In addition to viewing lessons from me, you'll have a chance to
contribute your own thoughts and ideas too, as part of the discussion forums, homework
assignments, and the peer grading process.
So that's the analytics part of this introduction to analytics modeling course, now let me say a
little bit about modeling. When most people think of modeling, this is the thing they think of.
Luckily for all of us, that's not what I'm going to be doing on these videos. You wouldn't want
to see it and I wouldn't want to do it.
In analytics, modeling means taking a real life situation and expressing the key parts of that
situation in terms of math. So we can analyze the math and then turn the mathematical
analysis back into a real life answer or solution or recommendation.
What is modeling :
In this short lesson, I'm going to try to head off a potential source of confusion. In analytics,
people use the word model to mean three slightly different things.
In the course introduction, we talked about how modeling means taking a real-life situation
and expressing the key parts of that situation in terms of math. So we can analyze the math
and then turn the mathematical analysis back into a real-life answer or solution or
recommendation.
The mathematical expression of the real-life situation is called a model.
2
,And confusingly, there are at least three different ways of using the word model and analytics. If
I say that I'm going to use regression to predict the delivery date of packages accompanying
ships out then regression is referred to as a model.
Anyway, if I plan to use regression to predict the delivery dates, then regression is the model
I'm using. And if I go one level of detail lower and say that I'm going to use regression to
predict
delivery dates based on a package's size, weight, and distance between where it's being
shipped from and where it's being shipped to then all that detail, the use of regression and the
use of size, weight, and distance is also called a model.
And if I go even one more step and use regression to find that the time for a package to be
delivered is approximately equal to 37 plus 81 times the sum of its linear dimensions plus
76 times its weight plus 4 times the delivery distance, then that specific equation is also called
a model.
Intro to Classification :
In this lesson, we'll talk about classification, including what it means, examples of when it's
important, and a simple demonstration of how classification works. Then in later lessons, we'll
see a basic model for solving classification problems.
Classification in analytics has the same meaning as it does in everyday life putting things into
categories.
The simplest examples of classification are when there are two categories, which are often
just yes and no. For example, a bank might want to differentiate between loan applications
who will
fully repay their loan and those who won't. A security agency might want to differentiate
between a regular person and a potential terrorist. An automated email filter might need to
differentiate between real and spam email. A legal document system might be designed to
differentiate between documents that are relevant or irrelevant to a certain case.
In my own research, we're working with the CDC on a project to increase the number of
organs that can be given to people in need of a transplant. And one of the questions we use
analytics to answer is, is this organ safe to transplant or does it carry a deadly infectious
disease?
3
, Of course, one answer would be to subject the organ to a laboratory test, but if the donor was
recently infected, the test might not come out positive yet. So we need the analytical approach.
In each case, we can use classification models to put applicants, people, documents, livers, etc,
into one of the categories. And as you can imagine, having more than two categories as well is
possible. For example, a political consultant might want to differentiate between supportive
voters, opposition voters, and undecided voters or a paleontologist might want to differentiate
between many different species of dinosaurs to determine which one in newfound bone belongs
to. Each of these classification questions will require some data in order to get answers. For
example, for loan applicants, a bank might collect data on income, credit history, age, family
size, assets, liabilities, and more.
Based on those attributes of previous loan recipients and the bank's observation of whether
each loan was repaid or not, the bank can then build a model to help classify future
applicants. Let's see how this would look graphically. Suppose the bank is trying to decide
whether or not to give loans to applicants using the applicants credit scores and incomes. So
we can plot each past recipients information on this graph. The horizontal axis will show
credit score and the vertical axis will show household income.
For example, someone with a credit score of 730 and an income of $76,000 a year would be
this data point. If they repaid the entire loan, we'll mark the data point in blue. If they defaulted,
we'll mark it in red. Of course, in real life, there might be other factors, each of which would be
another dimension, length of time as a customer of this bank might be a third factor, number of
dependents might be a fourth, total assets and liabilities would be a fifth and so on.
But as my wife, the artist, could tell you, I have enough trouble drawing clear two-dimensional
pictures. Three- dimensions is really iffy for me and I have no idea how to draw it on four more.
So we'll use this two-dimensional picture to see how classification works. As you can see, we
can draw a line like this that separates between the blue points above and the red points below.
So if we want to use this data to decide whether to offer a loan to a new applicant, we can just
see where the new applicants data point would be relative to the line.
An applicant up here, would be in the blue zone. So we might want to give them a loan.
An applicant down here, would be in the red zone so we might not want to give them a loan.
But how do we know that we've drawn the right line? There are lots of lines, infinitely many,
in fact, that would also separate the blue points from the red ones.
Choosing a Classifier :
In this lesson, we will see how to think about tradeoffs and classification problems graphically
to build your intuition before we get to the underlying mathematical models.
4
