Machine Learning Algorithms with R in Business Analytics

Professor Jessen Hobson & Ronald Guymon

Module 2: Framework for Machine Learning and Logistic

Regression

Table of Contents
Module 2: Framework for Machine Learning and Logistic Regression ...................................... 1
Module 2 Overview .......................................................................................................................... 2
Module 2 Introduction ..........................................................................................................................................2

Lesson 2-1: Machine Learning Overview ......................................................................................... 18

Lesson 2-1.1 Inference, Prediction, and Experimentation ..................................................................................18
Lesson 2-1.2 Categories of ML Models, Part 1, Types of Data and Terms ..........................................................30
Lesson 2-1.3 Categories of ML Models, Part 2, Categories of Algos ...................................................................39
Lesson 2-1.4 How Machine Learning Works in General .....................................................................................51
Lesson 2-1.5 Evaluating ML Model Quality .........................................................................................................60

Lesson 2-2: Business Problem ......................................................................................................... 70

Lesson 2-2.1 Introduce the Business Problem to Solve ......................................................................................70
Lesson 2-2.2 Introduce the Data .........................................................................................................................75

Lesson 2-3: Logistic Regression ....................................................................................................... 83

Lesson 2-3.1 Introduction to Logistic Regression ...............................................................................................83
Lesson 2-3.2 Logistic Regression Hands on - One Variable (Part1) .....................................................................93
Lesson 2-3.3 Logistic Regression Hands on - One Variable (Part2) ...................................................................103
Lesson 2-3.4 Logistic Regression Hands on - One Variable (Part3) ...................................................................117
Lesson 2-3.5 Logistic Regression Hands on - Multiple Variables ......................................................................131

Module 2 Review ......................................................................................................................... 140

Module 2 Conclusion ........................................................................................................................................140

1
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Module 2 Overview

Module 2 Introduction

In this module, we'll do two things. First, we'll introduce you to the world of machine
learning. Second, we'll examine a classic classification algorithm logistic regression.
More and more managers find themselves drowning in data. It's not just the everyday
transactions and sales data that stalks them. It's the shipping, the logistic data,
customer loyalty data, scanners and IoT data, that's machines talking to other
machines, supplier data, and employee, and HR data. The volume, velocity, variety of
data is only increasing if the use of the term big data is any indication. It's not just
hyperbole to say that we have more data available to analyze now than we ever have
before.

2
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Commentators state that over 2.5 quintillion bytes of data are created every single day.
That more than a megabyte of data is created every second for every person on Earth.
That means that in the time you've been sitting here listening to this video so far, more
data has been created just from you than most personal computers could handle in the
year 2000. Now think of your own employer. How much data is available to analyze?
How quickly does that data get created? Just keeping up with this data is a process in
and of itself. But you are tasked with not only understanding these mountains of data,
but also extracting actionable business insights from this data.

3
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Luckily, that's what this course is about. Helping you develop the ability to use tools to
extract the business insights from your data. Clearly, for most of us, there's too much
data to analyze by hand. Rather we need help. That's where machines come in.

This course deals with data modeling and using machines to gather insights from data.
After your data is acquired, cleansed, and ready for use, and explored, and understood
it at a general level, it's time to use statistics to draw actionable insight.

4
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

This module will discuss that general process, the menu of options available for
analyzing data, and general similarities that models have.

Machine learning has many definitions. But generally refers to solving a problem by
gathering data and then using a machine to follow an algorithm that builds a statistical
model based upon the data to gain actionable information from the data.

5
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

I put it in another way. A machine learning algorithm is like a program with instructions.
When you apply the algorithm to data, it creates a model. The model has new data
along with instructions for how to make predictions with the data.

6
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The learning part of machine learning refers to the fact that the algorithm can take the
data and use the statistics and parameters that you give it and arrive at information
without being explicitly programmed to arrive at that information every step of the way.
Thus, the computer takes the general directions you give it and finds information that
you did not explicitly tell it to find.

7
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Before we move forward, it's useful to discuss some additional terminology. Let's first
look at data science versus data analytics, versus business analytics. Unfortunately, for
those of you who like clear definitions and distinct differentiation between terms, you'll
be disappointed to learn that these terms are often used interchangeably. Most would
consider data science as the broadest term and both data analytics and business
analytics to be subsets of data science.

8
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Data science comprises methods and procedures to extract knowledge and insights
from data with the aid of computers.

Data analytics is similar, but a subset of data science that is more focused on describing
and visualizing data for specific objectives.

9
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Business analytics is a subset of data analytics that focuses data and questions related
to business and involves relatively few statistics.

Next, let's define and distinguish machine learning, artificial intelligence or AI, data
mining, and deep learning, etc. What do each of these terms mean and how are they
different?

10
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Well, all of these are disciplines in Data Science. Artificial intelligence is the broadest
term. It refers to the ability of machines to perform intelligent and cognitive tasks. Thus,
AI simulates thinking. Machine learning defined earlier, and data mining are both
subsets of AI.

Data mining is a cousin, if you will, to machine learning. It analyzes inputs to detect
outputs, but relies on direct human intuition rather than self-learning.

11
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Finally, deep learning is a direct subset of machine learning that is characterized by

using progressive layers of learning. That is, while many basic or shallow machine
learning models derive output directly from their input variables, deep learning produces
output based upon prior levels and prior layers of the model.

12
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, it's useful to understand where this relatively new field of learning, data analytics
and machine-learning came from. Data analytics is a confluence of many different
disciplines, including computer science, mathematics, statistics, information technology,
operations management, and business analytics. Thus, a data scientist and a data
oriented business analyst is a relatively new creation.

13
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

But perhaps at this point, it might be time to step back a little bit and ask, why we care
about machine learning at all? How does machine learning help us in business?

One way to think machine learning is just to relate it to the ways machines have helped
businesses throughout history. Certainly, no one would doubt the usefulness to
business of the steam engine or the telegraph. Like these early machines, computers
and business learning process data at phenomenal rates of speed and draw
connections, correlations, and insights that the analysts could get to know other way.

14
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Of course, don't be misled by the many movies predicting machines are taking over.

Machine learning helps the analyst derive useful and actionable insights from the data.

15
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

But it is still up to the analyst to act on these insights at least initially. Thus, machine
learning is like a great blood hound. It helps you gain insight, but must be directed
cautiously, or else it will run off chasing a squirrel. Finally, let's conclude by listing some
questions that machine learning can answer. Here are some categories of problems
that machine learning can help with and specific examples of problems from each
category. First, numerical prediction as seen in regression. What do we expect, for
example, costs to be next quarter in a segment? Next, cause and effect relationships
between variables. What what factors matter? Will this work? For example, which
factors have the biggest effect on shipping delay? Or does a particular change in our
website improve click-through traffic? Or does this layout of our store improve sales?

16
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, would be classification examples. What category does this observation belong in?
For example, is this transaction fraudulent? Or what factors increase the likelihood of a
purchase? Another category would be grouping things together. For example, which
group does this belong to? Which customers should receive which type of targeted
promotions? There's many others.

17
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Lesson 2-1: Machine Learning Overview

Lesson 2-1.1 Inference, Prediction, and Experimentation

Machine learning uses statistical models to gather information from your data. The
business analysts can then use that information to improve their business. There are
three primary categories into which we group machine learning insights. In other words,
machine learning helps us do the following three things. One, prediction and predictive
modeling, generally what will happen? Two inference, generally why does something
happen? Then three, experimentation. Generally, does changing something make a
difference? I'll speak of each of these categories in turn and describe some examples of
each.

18
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Prediction is one important use of machine learning. With prediction, you're asking
machine learning to look at the existing data and then to predict an outcome from new
input data. This prediction may be about a numerical value. For example, based upon
past data, how much do I predict sales will be next quarter? The prediction maybe about
which level of a class I expect new data to be from. For example, based on the data I
have, will this new information belong to a fraudulent or a non fraudulent class or
transaction? The prediction may also be about what group the data belongs to. For
example, based upon the data I've analyzed, a new data example or observation might
be from a customer who will likely become a loyalty customer versus one who will not.

19
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

If we are concerned only about prediction, we're not concerned about why the machine
learning model predicts a certain outcome, only with how accurate the machine learning
model is that prediction. Thus, we are indifferent about the underlying mechanism
generating the prediction. We don't care why it predicts accurately, we just want to
maximize accurate prediction. Thus, if we predict sales for the next quarter for a certain
retail store, we're not concerned with why sales will be a certain number, but rather only
whether the predicted number will be accurate.

20
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

But a second goal of machine learning is to understand why something happens or why
a certain outcome occurs. Under this paradigm, we're seeking to use the data we have
to make inferences about whether the variables we are examining will affect future
outcomes that we're interested in. Thus, we are concerned with which variables in our
data affect a target value of interest. Additionally, we are concerned with how these
variables affect this outcome variable.

21
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

We're seeking to develop stochastic models which fit the data and then to make
inferences about the data generating mechanism based on the structure of those
models. We're thus assuming that there's a true model generating the data. We're
concerned about the cause of the effects we see in the data. For example, instead of
predicting next quarter's sales, we might be interested in what caused that quarter's
sales and how our input variables affected those sales. Perhaps a percentage of our
sales that come from a certain product negatively affect sales. That is, the more of that
product that we sell, the lower sales actually are. This might happen if the sale of this
product reduces the sale of another more profitable product.

22
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Of course, frequently we care about both prediction and inference. That is, we care
about predicting future sales, for example, and we care about what variables affect
future sales. As another example, we might use machine learning to predict the selling
price of a house based upon that house's location, number of bedrooms, square
footage,

color of the front door,

23
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

lakeside location, etcetera.

In addition, we might want to know that the location of the house next to a loud highway
decreases the price of the house.

24
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The color of the front door does not significantly affect the price of the house

and the location on a lake increases the price of the house.

25
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The magic of statistics allow us to address these questions and more. Understanding
the ability of machine learning to make predictions and draw inferences helps us
understand the usefulness of machine learning.

Finally, these two categories highlight the need to use data to capture valuable business
insight. It's not enough to have data or even to understand our data. Our goal should be
to use our data to build better businesses.

26
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Thus, we should use tools like Machine Learning to experiment, be creative, and
explore. Experimentation is a final, important use of machine learning.

27
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

For example, a website retailer could test whether a certain change in the websites or
treatments increases or decreases traffic in that part of the website. Thus, the manager
leaves the website alone for some people, but then changes it for others and then uses
machine learning to determine whether this changed, increased or decreased traffic on
the website. This is often called A/B testing since the A group does not receive the
treatment, but the B group does. A is then compared to B using machine learning to see
if the treatment had a successful affect.

28
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Stephan Tom Key and others argue that companies should be actively building a culture
of experimentation. He believes that data trumps opinions and that opinions should be
supported by data. Again, machine learning makes it possible to experiment and to
know if those experiments succeed or fail.

In some, machine learning gives you the ability to predict, infer, and test and to let your
data speak.

29
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-1.2 Categories of ML Models, Part 1, Types of Data and Terms

Our goal in this video is to look at all of machine learning, the whole landscape and
build a foundation that will help you as you move forward and discuss individual
algorithms in more detail. To do this, we'll need to define some terms. Our first step in
building this foundation or this framework, is to understand important terms and
phrases. Thus, we will discuss terms used to describe the typical dataset used for
machine learning and define many aspects of it. Next, we'll introduce several different
categories of machine learning algorithms. Finally, we'll introduce some of the key
machine learning algorithms themselves. Let's start building our framework for machine
learning. In order to extract information from data, machine learning algorithms need
data, and data is why we're all here today. Broadly, there are two categories of data that
we'll talk about in machine learning; a structured data and unstructured data.

30
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

In this class, we're most concerned with structured data. Structured data is where each
example has the same features. These features are organized in a form that the
computer can understand easily. Thus, you can think of this as a spreadsheets or a
matrix.

31
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Each example is a row and each feature is a column. The second type of data is
unstructured data. These data are not structured into spreadsheets. They could be
sounds, pictures, and social media posts. In order to use these data will need to be
structured in some way.

32
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Again, in this class, we will focus on structured data in the form of a table or matrix or a
spreadsheet. Within a spreadsheet, there are two broad categories of data. One,
quantitative variables and two qualitative or categorical variables. Now a quantitative or
continuous variable is a number where each increments and the number higher or lower
has meaning. Further, the difference between two entries, for example, 10 and 12 has
meaning.

33
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

For example, you can think of revenue recorded for a product. A $10 product has
meaning and that is different from a $10.50 product. Also, that 50 cents difference
between those two has meaning as well. On the other hand, there's another category of
variable for which in-between measurements between two categories are not
meaningful. These variables are thus just categories, and we often call them categorical
variables or discrete variables. If there are just two categories like yes or no, true or
false, we'll call them dichotomous. A good example of this is a feature that lists cities in
the United States; Burlington and Waterloo. Two different cities, for example, are
categories that have no order. They're categories or classes because the difference
between these two means nothing.

34
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

That is half of Burlington, doesn't mean anything in this context, and R would store this
as a character. Now let's make one more important point. What if our column of cities
did not have names in it, like Burlington and Waterloo, but instead, replace those names
with a number with this column, then be categorical or continuous? This is an important
question because it changes how the machine it should treat and process that variable.
It's important that we get this right. It might be a bit confusing because this is clearly a
number, but the correct answer here is that this is a categorical variable still. The
numbers here are just arbitrary numbers that represent the cities and nothing else.
Thus, even though they're numbers, the difference between two cities, in this case two
numbers is still not meaningful.

35
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Now to be successful at machine learning, we need to speak the language of machine

learning. That is, we need to understand the terms that are routinely used. Thus, let's
continue to explore the terms of machine learning. Please be warned that different
people call these same things different names. Machine learning is not owned by any
one academic discipline. It's part of statistics, part of computer science, and a bit of lots
of other disciplines. Thus, the different disciplines give these things different names. I'll
do my best to give you the most common names. A unit of observation is the smallest
entity with measured properties of interests for study. Thus, for example, one product
sold in a transaction could be a unit of observation. Next, a unit of analysis is the
smallest unit from which inference is made. Thus, while one product from a transaction
may be the smallest unit of observation, you may be more interested in aggregated
sales over the course of a month. Thus, the unit of analysis might be the monthly sales
in a product category.

36
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, examples, cases, values, and observations are instances of the unit of
observation. In our case, since we're using structured data, this is a matrix or a
spreadsheet, and each example is a row. Next, features, variables, dimensions,
attributes, inputs, and predictors are properties or attributes of examples. Again, in our
case, using matrix data, each feature would be a column. For example, if we're
examining products sold in individual transactions, the feature could be the product
itself, time of sale, location of sale, whether a customer had a loyalty card and etc.

37
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Frequently in machine learning, we're examining the effect of multiple feature of

variables on one target variable. We want to predict that target variable or examine how
the other features affect the target feature. When this is the case, the target feature can
be called the dependent variable, the output variable, the sponsor variable or y. The
other features will then be called the independent variables, the input variables or x. For
example, we might want to predict next quarter's revenue for a store that we manage.
Thus, the target or dependent variable is next quarter's revenue and the independent
variables or the input variables, could be current revenue, direct costs, costs of goods
sold, seasonal demand, etc.

38
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-1.3 Categories of ML Models, Part 2, Categories of Algos

Next, in our effort to build a framework for understanding machine learning, let's explore the
many different types of learning algorithms that exist. It would be really convenient and this
class would be much shorter if there were one algorithm that would work in every situation.
Unfortunately, every situation for which we can use machine learning is different since no two
datasets are the same. Thus, as D. H. Wolpert says, there's no free lunch in machine learning.
There is no one model that works best for all possible situations. There are generally three
distinct categories of machine learning algorithms:

39
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

supervised, unsupervised, and reinforcement. In addition to these three main categories, we'll
also talk about semi-supervised learning and also data mining.

Supervised learning then is characterized as having a labeled dataset. That means that all
features, including the feature of interest are labeled, and that is, they all have a value. Thus, a
supervised learning algorithm uses data to create a model that finds a combination of features
that results in a target feature or label, which can be a number or a category or class.

40
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's state this in a slightly different and maybe more helpful way. A supervised learning
algorithm uses a model to analyze and decipher the relationship between features and an
existing target feature. For example, regression is a supervised learning algorithm. They could
use features such as revenue last quarter, revenue last year, forecasted weather, etc, to predict
the target feature of revenue for the next quarter. Next, a classification algorithm is a supervised
algorithm that could use features such as time of day, amount of purchase, location of
purchase, etc, to predict a label for the target variable of whether, for example, a purchase is
fraudulent or not. Unsupervised learning is characterized by a dataset of unlabeled examples,
that is, features without a target feature. The unsupervised algorithm transforms the features
either into another vector as in dimensionality reduction, or into a value as in clustering.

41
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's put this into different, perhaps easier terms. Since only input variables are present and no
targets or output variable exist, these algorithms analyzes relationships between input variables
and uncover hidden patterns that can be extracted to create new labels for output variables. For
example, in a clustering algorithm, you take features such as an item purchased, time of a
purchase, etc, to cluster customers into groups. For example, the model might determine, based
upon convenience store purchases, that a customer belongs to a fuel group, or a fountain soda
group, or a weekend alcohol group, or a daily lottery group.

42
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, reinforcement learning uses data to create a policy that takes features of the data and
outputs an optimal action for that state. For example, think of learning to play a video game.

43
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

You want to maximize the outcome of success and at each stage, you learn more and use that
learning to reinforce and influence subsequent gameplay. For example, reinforcement learning
can help a car learn not to crash and can help a computer learning from scratch how to play
chess. Next, semi-supervised is just like supervised learning, but some of the target variables
are missing. Thus, you're training a model with labeled and unlabeled data. This can happen
when labeling the data manually all the way through. It's too costly. Thus, you could build a
model with labeled cases and then use that same model to label the remaining cases.

44
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Finally, data mining is frequently used to refer to different machine learning techniques. Thus,
many sources see data mining as the very same thing as machine learning. If we're trying to
distinguish data mining from machine learning, we would say that machine learning analyzes
inputs to detect outputs, but relies more on direct human intuition than on self-learning.

45
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Finally, let's mention just a few of the key algorithms in supervised and unsupervised learning.
Supervised learning uses regression to predict and examine target variables that are numerical
or continuous. An example of a regression problem, that is, a business problem for which a
regression algorithm could provide insight, is investigating the quarterly sales by store of a set of
convenience stores.

46
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

A regression model could use input variables such as store location, past sales, and current
costs to predict future sales. This model would allow the analyst to examine the effect of these
input variables on the output variable, quarterly sales.

47
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

It would also be useful to predict next quarter sales, which is very valuable for forecasting and
resource planning. Notice that the target feature in regression is a continuous variable. On the
other hand, classification algorithms such as logistic regression, decision trees, random forests,
and k-nearest neighbors, predict and examine target variables that are categorical.

They create decision boundaries that separate target variables into different classes.

48
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

An example of a classification problem is a company that wants to predict whether a purchase

will be from a loyalty customer or a non-loyalty customer. A logistic regression algorithm could
use a model to take the input independent variables and predict whether the output dependent
variable is a loyalty customer or not. This model would be helpful because it can be used to
predict what type of product is often purchased by a loyalty customer, then that product could be
used to promote the loyalty program, particularly in locations that have poor loyalty system
adoption. Next, in unsupervised learning, clustering, k-means clustering, density-based
clustering, group observations together based upon their features. Clustering algorithms, unlike
regression and classification, work on only input variables, often called unlabeled data, to put
different observations into groups based upon those variables.

49
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

A clustering problem could be that a convenience store business wants to categorize its
customers into different groups, such as a gasoline group, a morning coffee and breakfast
group, and the lunch group. Being able to identify customers in this way could lead to targeted
promotions and discounts to increase traffic and increase the size of purchases of each cluster.

Here's a final graphic that includes a sample of the whole landscape of machine learning.

50
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-1.4 How Machine Learning Works in General

In this video, we'll discuss the basic level, how machine learning algorithms and models work.
Now it's important to keep in mind that although there are many different machine learning
models, options for those models called hyperparameters and questions that can be addressed
with machine learning, all machine learning algorithms follow the same basic process. They
take data and identify patterns that form the basis for further action. There's six general steps
necessary to use machine learning. These are the following: one, data scrubbing, two, algorithm
selection, three, model training or fitting, four, model evaluation, five, model improvement, and
six, deploy the model. I'll talk about each of these in turn.

51
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The first step in running a machine learning algorithm is getting, loading, transforming, and
cleaning data. After the initial ETL or extract, transform and load process is complete, the data
is in a usable form and relatively free of mistakes and problems. There are usually a few more
data cleaning issues that might be necessary based upon the particular algorithm that will be
used. For example, redundant columns might need to be eliminated. For example, a data set
that lists products and has one column for the product name, and another column with a unique
numerical identifier for each product has a redundancy. In this case, one of the columns should
be removed. Eliminating and selecting columns is often called feature selection. Another data
scrubbing activity would be to replace text in categorical variables with numbers. This is
sometimes called one-hot encoding. For example, a column that lists cats and dogs could be
changed to a column that lists zero for cats and one for dogs. This is also called dummy coding.

52
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Another data preprocessing technique that may be beneficial given the machine learning
algorithm that you use is standardization. Standardization, it takes variables that are on different
scales and puts them on the same scale so that they have the same variance. For example,
assume your data set list financial statement numbers for a company, and that's some of the
numbers are in thousands while others are in millions, standardizing these figures by
subtracting the mean and dividing by the standard deviation would give both measures a mean
of zero and a standard deviation of one. Standardizing in this way is generally recommended for
principal components, support vector machines, k-nearest neighbors, and other algorithms that
use distance measures. A final topic of interest is not how the data should be scrubbed but how
much data is needed in the first place. Well, there's no single definitive rule here, there are rules
of thumb. Ideally, you have at least some observations for each combination of levels of the
variables. For example, if a data set has two categorical variables with two levels each, for
example, animal with cat and dog and age, old or young, it would be ideal to have data for old
cats, young cats, old dogs, and young dogs. This way, things won't unravel at the side of
unseen combinations in trying to predict the data in the future. A rule of thumb for the minimum
amount of data to have to run machine learning algorithms is 10 times as many data points as
the number of features. Ultimately, how much data is required depends on the algorithm that's
being used. Clustering and dimensionality reduction algorithms can be effective with less than
10,000 observations. Regression and classification algorithms can get by with less than 100,000
observations while neural networks need more than 100,000 observations.

53
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The next step is algorithm selection. Selecting an algorithm that's best for your analysis project
is one of the business analysts hardest jobs. The analyst examines their question, their data set,
and they need a solution and picks accordingly. Unfortunately, there's no one model that's best
for every business problem and business data set. Analysts might be tempted to reuse models
that they're most comfortable with, but the analyst should be aware of all the models available
and also needs to be aware of recent advances in the field.

54
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The next step then is to apply the algorithm to the data to create a model. The model takes the
data set, following the rules of the algorithm, and converts it into information that can be used to
solve a business problem. That is, the machine learning model includes new data generated
from applying the algorithm to the data set, as well as instructions for the machine for how to
use that data to create predictions on new data. Frequently, but depending on the algorithm, the
model is created on a subset of the data called the training data set. Specifically, the data set is
split into two separate parts, called the training data and the testing data. For example, the data
set might be split 70 percent for the training and 30 percent for testing. A 60, 40 or a 80, 20 split
could also be used. Now this is sometimes called split validation. Once the model is trained on
the training data, the testing data can be used to check the accuracy of the model. Thus, the
point of splitting the data is to check whether the model that is created works not just on the
data that we have, the training data, but also on the new data, the holdout data, or the testing
data as well. This avoids the situation in which our model fits the data we have perfectly, but
does not generalize to the other data. That is, it does not work for new data. This is important
because the goal of many machine learning solutions is to speak not only to the data we have,
but also to predict future outcomes and future relationships between variables.

55
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

After the model is trained, the next step is to evaluate how accurate the model is. As just
mentioned, the goal of creating the model is to explain or predict the data we have, but also to
explain and predict new data. Since our true goal and training a model is to solve a business
problem, we want our model to predict the truth as closely as possible.

56
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

To evaluate the model's ability to do that, we examine measures of accuracy for the model. For
example, if our machine learning task is to accurately classify whether a purchase is fraudulent
or not a classification task, we would examine how accurate the model is at predicting fraud
within our training data set. Next, we would examine how accurate the model is at predicting
fraud and the testing data set. If accuracy is relatively high for both the testing and the training
data sets, we'll have some assurance that the model will predict fraud well for new data.
Measures of accuracy for classification models, include the area under the curve, AUC, the
receiving operating characteristic, the ROC, a confusion matrix, recall and etc.

57
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

If the accuracy of the model is not good enough, the next step is to improve the model. Many
methods exist to improve the quality of the model. In fact, competitions to create the most
accurate and useful model are a significant part of the machine learning culture. A first step for
seeking to improve the accuracy of the model is to change the instructions given to the
algorithm. Algorithms have options that control and impact how the resulting model learns and
which patterns to identify and analyze. These options are called hyperparameters. These
options can be changed by the analysts to fine tune the model to better fit the data set. Although
fitting the model too closely to the data set can cause a problem called overfitting. Next, the
analysts might increase the amount of data. All data has noise or irreducible error caused by
missing variables or a unique, idiosyncratic variation. Increasing the amount of data can reduce
that variation. Similarly, feature selection, that is adding additional variables that strongly affect
the target variable and removing less useful variables may help. Thus, it's important to
understand that the data scrubbing and cleansing process can sometimes just beyond going.
Finally, the analyst might need to pick an entirely new algorithm.

58
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The final step in this process is to deploy the model and put it into production. The whole point
of using machine learning in the first place is to create a model from your data that helps you
address business problems. An accurate model can give you support to implement and test a
change in your business, reallocate resources to create or emphasize different classes of
products or customers, forecast future financial results for a segment of your business, and
classify new data as helpful or problematic.

59
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-1.5 Evaluating ML Model Quality

In this video, we'll discuss the important topic of machine learning model accuracy. Now, this is
a nuanced topic with no simple, easy solution. The business analyst has to make trade-offs, and
it's important for us to spend a little time to understand the issues at hand.

60
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

At first, it's important to remember the goal of using machine learning in the first place. Though
machine learning can be used for a lot of reasons, at its core, we're interested in creating a
model based upon our data that predicts or uncovers truth of some kind. Once that truth is
found, we can use it to make better business decisions and solve business problems. For
example, we might want to know the truth about how a certain store will perform next quarter.
Alternatively, we might want to know the truth about whether a new purchase is a real purchase
or a fraudulent purchase.

61
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The problem, of course, is that it's difficult to know the truth. Now, we're not trying to get
into philosophy here. We're just stating the fact that the dataset we're using to find this
truth is incomplete. That is, all data has irreducible noise. Unless you're looking at every
single data point that you care about, some, hopefully, a lot of your data's
characteristics and relationships will translate to other similar data. But there will always
be some portion that does not translate. This could be because you don't have all of the
important features of the real world in your data, or just because all the data is a little bit
unique. For example, say you have a dataset that lists transactions and whether those
transactions are fraudulent or not. If your business problem is to predict whether a
transaction is fraudulent or not, your reason for building a model is not to perfectly fit
your current data, but rather to use that information to predict whether a future
transaction is fraudulent. Your current data can help with that, but it does not contain
every possible transaction of every variable that would explain whether a transaction is
fraudulent. Further, the data you collected last Tuesday, for example, will be unique, at
least in part to last Tuesday, and will not be 100 percent representative of every
Tuesday ever in the future. Furthermore, there might be small errors or mistakes in that
data.

62
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The key point is that because all data has some noise or irreducible error in it, if you
want your machine learning to apply to more than just this data, you need to make an
intelligent trade-off between your model being accurate, but not too accurate. Because if
you're model matches your data too perfectly, it will likely not match new data very well.
This is what we call overfitting. Thus, there's always a trade-off between training a
model that's good, but not too perfect, not too good. You want your model to generalize,
but not to memorize.

63
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

How does this work in practice? Suppose you have some data and you want to use a
machine-learning algorithm to train a model that will predict a future data point. One way
to avoid overfitting is to split your data into testing and training segments called split
validation. Thus, you split your data into two groups of 80 percent and 20 percent, or 70
percent and 30 percent.

64
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

A key step to remember is certainly to randomly sort your data before splitting it into
your training and testing sets. Most datasets come sorted in some logical order. If you
keep the data in this order, it's likely that your training set will be systematically different
from your testing set since they're on different ends of meaningfully ordered data. For
example, the data might be sorted alphabetically based upon date. Thus, the testing
data would be systematically newer than the training data.

65
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, to help make sure your model will generalize to new data, you train your model on
the bigger portion, the training data, and then test the model on the smaller portion, the
testing data. Then you compare the accuracy of your model in the training data to that in
the testing data,

you would like the accuracy of the model to be high in the training data. If it's not, we
call that underfitting. If the accuracy is low in the training data,

66
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

you'll need to add complexity to your model or pick a new algorithm with a more flexible
model. This will help the model match the data points more specifically.

A model that underfits the training data is said to have high bias.

67
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, you run the model on the test data. You want the accuracy of the model to also be
high in the test data. If it's not high, you have overfit your model to the training data. The
model is too specific and does not generalize. This means the model is responding too
much to random chance or noise that's in the training data. If this is the case, you can
add more data to the training set and/or reduce the complexity of your model, or pick a
new algorithm with a less flexible model. A model that overfits the training data is said to
have high variance.

68
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So our constant effort as analysts is to pick algorithms and tune models to avoid
underfitting and overfitting, that have low bias and low variance. Let's briefly summarize
this discussion by using this picture. Suppose the dots here represent your training data
and suppose that the three lines are the predictions of three separate machine learning
models, the red line is too simple and biased and underfits the data. However, the blue
model appears to go too far. It's too complex and specific, it has too much variance and
overfits the data hitting every point exactly. It's likely that when this model is run on the
testing data, it will not fit the testing data well, because it so specifically matches the
training data. Finally, the green model seems just right. It's more complex and specific
than the red model, and thus does have some more variance, but it does match the
data well, and thus reduces bias. We'll have to check with the testing data, but it
appears that the green model avoids both overfitting and underfitting and will thus
generalize well.

69
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Lesson 2-2: Business Problem

Lesson 2-2.1 Introduce the Business Problem to Solve

In this video, we'll set up a realistic hypothetical business problem. In subsequent

videos, we'll use a new data analytics tool to gain useful insight into this problem. The
hypothetical company will be helping is TECA. TECA owns over a 150 convenience
stores and gas stations throughout the middle of the country. TECA knows from prior
research and general industry knowledge, that a new loyalty customer is worth about
$200 earn extra revenue a year. A loyalty customer is an individual who has provided
certain information to TECA convenience store in order to receive a loyalty card. Loyalty
cards provide discounts and benefits for those who use them frequently. You probably
have a loyalty card from your favorite grocery store, convenience store, ice cream shop,
or movie theater. These businesses know that the small discounts they give you on your
loyalty card are more than compensated for by the increase in business you provide to
them.

70
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Not only do loyalty customers spend more money, as can be seen in this graph, loyalty
customers also buy higher margin products. This graph shows the gross margin profit
for non loyalty and loyalty customers. You can see that the loyalty column looks
significantly higher than the non loyalty column. Thus, loyalty customers not only come
into the store more often, they also buy more profitable items when they do come.

71
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

TECA would like to be able to predict when a transaction is more likely to be made by a
loyalty customer for two reasons. First, they'd like to increase the percentage of their
customers who are loyalty customers. Second, they would like to increase the number
of transactions those customers make. Thus, TECA is embarking on a research project
to examine how they can make transactions more appealing to loyalty customers.

72
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

What TECA would like to do is predict when a transaction will be from a loyalty
customer. Perhaps certain products drive loyalty purchases. If that's the case, TECA
can promote these products to non loyalty customers to encourage participation in the
loyalty program. Perhaps there are certain times of the year when loyalty purchases are
more likely to occur. If so, TECA could advertise accordingly by focusing on promotional
programs on those months in which loyalty is most or least likely to occur. Finally,
particular regions of the country might be better at selling to loyalty customers. If those
areas and stores are identified, TECA could investigate their underlying success and
replicate it elsewhere to make sure that more purchases are made by loyalty customers.

73
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Thus, TECA would like to build a model that predicts when a transaction is likely to be
from a loyalty customer versus a non loyalty customer. In addition, TECA would like to
explore how different products, store location, and the time of year, affect that model.

74
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-2.2 Introduce the Data

In this video, we'll take a brief look at the data we'll be analyzing throughout this module.
We'll be using a neat data set created for this course called the TECA dataset.

TECA is a fictional company that owns over 150 convenience stores and gas stations
throughout the middle of the country; from Wyoming to Minnesota, from Arkansas to
Alabama and many states in between.

75
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The particular dataset we'll be using is a random selection of customer transactions

from these convenience stores for the years 2017-2019. These transactions include, for
example, purchases of fuel, lottery tickets, soda and candy. Though our dataset is
relatively large, including over a million transactions, it only includes a pretty small
random sample from all of the transactions in this period. The full dataset would be over
a 145 times larger than this and consist of hundreds of millions of transactions.

76
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

It's important to know the level of aggregation of the data. That is, what each row
represents. In this case, each row represents a single product purchased as part of a
single transaction. Thus, the data is organized at the transaction product level. Each
row has an entry in the column unique ID. This number is unique to that row, meaning
that no other row has that number.

77
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Additionally, each row has an entry for transaction ID. This number is not necessarily
unique since one transaction can have multiple products. For example, unique ID
equals 15 could represent the purchase of an extra large fountain drink as part of
transaction ID 20189999434333452532. While most transactions are for only one
product, it is possible that this same transaction also included the purchase of an apple.
Of course, the apple would be unique ID equals 16, since each product for each
transaction has its own unique ID.

78
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

There's dozens of features or columns in the data. We won't describe them all here, but
they can be grouped into five categories. We'll describe the five categories and the main
columns or features in each category.

First, the first group of columns is the transaction group. These columns include the
unique identifier for the row, the transaction number, and the date of the transaction.

79
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Second, the next group of columns is the customer group. Actually this is not a group of
columns, but only one column called Customer ID. This column identifies the customer
making the purchase when that customer is known. Customers are only known when
they are loyalty reward members of the company and made their purchase with a loyalty
rewards number or card.

80
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The third group is the product and category group. This group of columns identifies the
product and its name. It also identifies the type of category of product. All products are a
part of a hierarchy of product categories. Thus, each product belongs to a category and
that category has a parent category or a superset of that category.

81
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, the fourth group is the site group. These columns identify the convenience store
where the purchase took place, including the site name, the address of the site, and the
longitude and latitude. Please know, however, that to obscure the actual company using
this dataset, the addresses are masked by referring to post offices instead of the actual
convenience stores themselves. Nevertheless, the addresses are close to the
convenience store. Finally, five, the last group of columns, present the amount of
revenue and profit for the product that was sold. These columns are critical to all of our
analysis going forward since TECA is a business and is interested in making money.

82
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Lesson 2-3: Logistic Regression

Lesson 2-3.1 Introduction to Logistic Regression

Our hypothetical company TECA wants to be able to predict whether an individual

purchase is going to be made by a loyalty customer or a regular non loyalty customer.
Loyalty customers are those who applied for a loyalty card with TECA in order to get
discounts for purchases. TECA a has learned that they make about two $100 per new
loyalty customer, and that loyalty customers buy more profitable products on average
than non loyalty customers do. By predicting whether a sale is made by a loyalty or non
loyalty customer take a hopes to increase the number of loyalty customers and the
number of products that they buy.

83
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

While there may be many ways to investigate this issue, TECA has approached the
problem by labeling each of their purchases as being made by a loyalty customer
labeled as loyal or made by a non loyalty customer labeled as not loyal. Thus the
targets or dependent variable here is loyalty and it's a categorical variable.

84
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The other features, or independent variables that will be used to predict the target
feature are the following category. This variable list, the top 20 most frequently
purchased products and they are the following fuel, which is gasoline purchased at the
gas pumps, pop, which is bottled and canned soda, cold dispensed beverage, which is
fountain soda. Next juice and tonics, these are bottled juices,

85
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

lottery, these are lottery tickets, both those purchased and paid out if it's a winning
ticket. Energy, these are bottled and canned energy drinks. Next candy and gum just
self explanatory, hot dispense to beverage, these is coffee, tea and hot chocolate from
a dispenser.

Next beer and then salty snacks, these include peanuts and sunflower seeds, for
example. Cigarettes, pizza,

86
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

roller grill, these items are hot dogs and burritos that are kept warm on a roller grill.
Breads and cakes, breakfast sandwiches, these refreshing take and heat breakfast
sandwiches. Then bakery, these are donuts and cupcakes.

Another one is smokeless these, this is chewing tobacco and then hot sandwiches and
chicken, these are ready to purchase sandwiches and chicken cigars and then chips.
Here's a snapshot of the data that you can see.

87
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

This data has been significantly cleans and manipulated for use in this module, null
values and outliers have been removed and categories have been combined and
eliminated. Finally, the data has been adjusted so that the number of loyalty purchase
and non loyalty purchases are relatively similar.

88
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Now take a wants to predict loyalty based upon the three variables, category, Quarterl,
and state. This feels like a solution that regression could handle, certainly regression
examines the effect of independent variables on dependent variables and predicts an
outcome. Unfortunately there is one major difference between a regression problem and
our problem, that is regression predicts a continuous dependent variable. While we
need to predict a categorical variable

89
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

specifically, as can be seen here, we have by variant data that is all of our data lies at
two levels loyal and not loyal. Regression would pick a regression trend line that gives a
continuous response and would fall outside of the range of our dependent measure.
Rather than that we need a function that converts the independent variables into an
expression of probability between zero and one in relation to the dependent variable
which would be not loyal and loyal. That is we need a function that gives outputs
between zero and one that finds coefficients to maximize the likelihood of predicting a
high probability for observations that actually belong to class one, which in our case is
loyalty. And predicting a low probability for observations that actually belong to class
zero, which in our case is not loyal. This will then produce an S shaped curve that can
convert any number and map it into a numerical value between zero and 1.

90
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Now luckily there are many functions that can do this for logistic aggression will pick the
logistic function or the sigmoid function. Here's a picture of a multi variant version of
that,

this function provides a nice S curve that gives us precisely what we're looking for. Thus
this function models the probability that why the dependent variable belongs to a
particular category given the x i the independent variables.

91
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

In other words, based upon the found probabilities of the independent variables, logistic
regression assigns each data point to a discrete class, in our case, loyal and not loyal.
Unlike the ridge regression prediction trendline, the logistic hyper plane represents a
classification decision boundary that divides the data set into classes. Finally, the
estimation of coefficients of this model is done via maximum likelihood. In general terms
you use maximum likelihood to find the intercept and coefficient of X such that the
predicted probability of Y is as close to possible to 1 for true y equals 1, and zero for
true y equals zero. In summary, as it's probably clear by now, logistic regression is a
classification machine learning algorithm. It's useful for prediction and exploration of
categorical target variables. It's just one of many different classification algorithms.

92
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-3.2 Logistic Regression Hands on - One Variable (Part1)

In this video, we'll use the algorithm to predict whether a transaction is from a loyalty customer
or not. We'll use the TEKA data set. TEKA desires to increase their loyalty customers and the
number of products a loyalty customer buys because loyalty customers create more business,
and more profitable business for TEKA. As always, we encourage you to resist the urge to just
watch these videos. Rather, please follow along doing each of the steps that we're doing so that
you can get the critical hands-on experience necessary to learn this valuable tool. First, let's go
ahead and bring in the data-sets and the needed packages.

93
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

I'll do that here. The package we're going to use is tidy verse. We put that in the library function.
If you don't have the tidy verse package, you'll need to download that yourself. Then we bring in
a file that we'll call df1_even_kp. We'll use the read_rds function and the rds file that we provide
for you as logistics one.

Next, let's run the following lines of code. In these lines, I've created a very basic function that
will help us evaluate the quality of our logistic regression model later on. We'll use this code
several times so it makes sense to put it into a function that can be re-used without having to cut
and paste all the code every time that we want to use it.

94
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

I won't describe this here, but you'll need to go ahead and run it.

95
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, let's go through our six steps for using a machine learning algorithm. Recall that these
steps are the following: one, data scrubbing; two, algorithms selection; three, model training or
model fitting; four, model evaluation; and five, model improvement with; six, deploying the
model. We've done all the data scrubbing already for you so number 1 is already complete. For
number 2, we selected logistic regression as our classification algorithm so let's move on now to
Number 3. We'll break this down into multiple parts. Our goal is to train a univariate logistic
regression model using our data. That is, we'll predict whether a particular transaction is from a
loyalty customer or not.

96
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's look at the data to remind ourselves what it looks like. To do that, let's run this
slice_sample function. Just to get a sense of what the data is, we'll pass n equals to 10 to the
function so that we can view and visualize just the first 10 rows. Loyalty, this column right here,
this first column, that's our dependence or output or target variable. I'll use those terms
interchangeably. Then we need to pick one of the other three, category, quarter, or state, to be
the independent variable.

97
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Since TEKA is concerned with predicting what transactions will be loyalty transactions, it
probably makes the most sense to use category as our independent variable. Now, before we
train our model, we have a few things to do to get ready. First, let's look at the baseline
percentage of the occurrence of loyalty. We'll use the function table to do that. This function
uses the levels of a categorical variable to build a contingency table of the counts at each
combination of variable levels. We'd like to build a table of how many loyal and non loyal
transactions that we have and then calculate a percentage of loyal transactions. Let's describe
what's being done here. First, the table function is used to create a table from the loyalty
variable. Second, we print the table. Next, we print the percentage of loyal transactions divided
by the total transactions. This is done by indexing the table and using indexing to select
individual parts of the table with the square brackets. That's for example, loyalty Table 2 is
519744.

98
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Again, we create a new table called loyal table. We take the table function and we pass the one
column from df_even_kp, which is the loyalty column, then we print that table, and we simply
index, by using the square brackets, the different parts of that table to create our percentages.
There we go.

99
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Overall, loyal transactions happen about 49, almost 50 percent of the time, as you can see from
this number here. Next, I want to double-check that the levels of loyalty are in the order I think
they are. Logistic regression and the confusion matrix that I created, that first thing we ran way
before, is going to use factors in the order that they're in, in that variable. I need to understand
what the correct order is so that I can interpret the output of the model correctly. The contrast
function will check that for me. I'd like to have loyal be coded as one.

100
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's go ahead and run that. Again, I'm simply passing the one column, loyalty column, in that
contrast table. Sure enough, that's correct. I can see that zero is not loyal and one is equal to
loyal. We can move on to our next step.

The next step in training our data is to split it into training and testing proportions. Recall that we
want to train our model on one set of data but then make sure it works on the other data so that
we're not over feeding our model. I'll use the caret package to split the data though there are a
million of ways of doing this. However, as you continue to explore and work on classification
models, it is likely that you'll probably use the caret package in the future.

101
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

This first line here just brings in the caret package using the library function. This next line uses
set.seed and then just a number I put in there, 77, to create a seed so that in the third line, the
create data partition is going to use the same seed or same starting point every time so we get
the same split of data every time. This line then uses the caret package to select 75 percent of
the entries in the column loyalty. We could have used a different percentage. We could have
used 80 percent, we could have used 70 percent. Let's go ahead and go with 75. This, then
creates a matrix of numbers that we can use to select the rows from our data set that will
become the training data. That is what we've done in the fourth line here. Here, we create the
data set called data train by selecting the original data set and telling it to only keep the rows we
randomly selected in the partition that we created right above. The square brackets here are
again using indexing to tell R which rows to take, which rows to use. We put nothing after the
comma right here which tells R to take all of the columns. Just the rows from partition, but all of
the columns. We could have instead selected only some subset of the columns. Next, we select
the opposite columns in data test using this minus or dashed symbol in front of partition to make
the testing data set, data_test. Finally, we print the percentage of rows that are in the training
data just to make sure that it's really close to 75 percent as we intended. We run that, and sure
enough we have 75 percent here.

102
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-3.3 Logistic Regression Hands on - One Variable (Part2)

With our Data Split, we're ready to train our model. Recall that an algorithm is a set of rules for
how to make a model. A model takes our data and the options we select to create new data,
new information, and additional rules for how to use that model. Below we create our model
called model train. Then summarize it to view the output, we use the GLM function with the
family equals binomial argument to select logistic regression model. The syntax is quite simple
here, our dependent variable, loyalty goes first. Then we use the tilde sign, that squiggly line, to
act like an equal sign, followed by our independent variable, category in this case. Finally, we
list the data set. We are of course using the training data since we're training the model so far.

103
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's go ahead and run that and we'll also summarize it with the summary function. I'll take a
moment to train that and you can see we're running it and there it is.

There's a lot of data here and let's pick out only the most important things and talk about those.

104
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

We'll focus first on the coefficients area. There we have coefficients listed for our variable under
the estimate column, right here. Here's the coefficients for each of those category levels.

105
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Away over to the right here we have the z-values and the p-values, here and here. What does
all this mean? The coefficients tells us the effect of that particular variable or level of variable on
the dependent variable which is loyalty. The GLM function is smart enough to know that
category is a categorical variable and not a continuous variable. Thus, the function automatically
did something called one-hot encoding. Equivalently, we could call that dummy coding. It did
that to category all on its own. That is, it took the variable and essentially created individual
dummy variables for each level of category. You can see those here. Thus, instead of having
one category variable, we now have 19 dummy variables for each level of category or almost
each level. Thus, category beer is a new variable that has a one if that row is as a beer
purchase and a zero otherwise. That's an example of a dummy variable. Likewise, category
pizza, for example, has a one if that level row has pizza in the category column and a 0
otherwise.

106
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Now hopefully you noticed an issue. We have 20 levels of category, if you remember back from
above, but we only have 19 dummy variables here. That's because GLM smartly puts one of the
levels of our category variable in the intercept term, the intercept right here.

107
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Specifically, bakery is in the intercept. This is done to avoid over specifying the model. It's
important to know which level is not in the model and which is in the intercept because it affects
how we interpret the results. Specifically, the effect of the other coefficients is in reference to the
variable that's in the intercept which in this case is bakery. We can now interpret. The coefficient
on intercept here lists the effect of bakery items on whether a transaction is from a loyalty
customer or not.

108
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The coefficient here is positive, which means that selling a bakery item increases the likelihood
that a transaction is a loyalty transaction.

More precisely, purchasing this item increases the log odds of a loyalty purchase happening by
0.30575. Now we won't go into log odds, but it suffices to say that buying a bakery item
increases the likelihood of a transaction being from a loyalty customer. Let's look at some more
of our dummy variable coefficients here.

109
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Let's look at fuel for example. Fuel actually decreases the chance of the transaction being from
a loyalty customer relative to a bakery item. Again, bakery is in the intercept,

110
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

as can be seen by this negative coefficient here. More specifically, the coefficient is still negative
even relative to the intercept, which again is bakery. That is, fuels coefficient plus the intercept
is still negative if we add those two together. This means that fuel decreases the chance that a
loyalty customer makes a purchase. Now this does make sense since bakery item is one that
you go into the store for, which is more likely to happen for a loyalty customer, whereas anyone
can just drive by and purchase gas. Next, let's look at fountain soda or cold dispensed
beverage. This increases the chance the purchase was from a loyalty customer, even more
than purchasing a bakery item. Thus, fountain soda seems to be a big draw for our loyalty
customers and it could be a way that take I could bring increased loyalty purchases. Again, we
noticed that specifically by the positive coefficient,

111
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

the fact that the 0.40645 is even larger than the 0.30575. Finally, let's look at the effect of bread
and cakes. What effect does this have? It's a small negative coefficient,

112
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

but if we add this negative coefficient, the negative 0.11000, to the intercepts, we see that the
difference is actually positive. Again, adding those two together. Thus, bread and cakes has a
significant effect on the chance of purchase as a loyalty purchase. Again, we have to compare
with the intercept because the bakeries in the intercept and we don't want to over specify our
model. Next, it's not enough to see what the sign of the coefficients are that's what we've been
doing. But we need to see whether the coefficient has a statistically significant effect on loyalty,
we see this in at least two ways. First, we can look at the p level.

113
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The p are greater than z right here, this last column or next to last column. This is the probability
that the coefficient we're particularly looking at is equal to zero.

This is not the technical definition for that but it's a point to remember. The thing that you should
retain is that the lower the p-value, that means that the coefficient is not equal to zero, and thus
is statistically significant. Now, additionally, just to the right of that,

114
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

you see these asterix here, these little stars. These also indicate the significance of the
coefficient.

115
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The more stars, the more significant it is, the more likely that the coefficient is not zero. We have
a lot of data here and even small coefficients are likely to be significant. Sure enough, only
energy drinks if we look down, and hot sandwiches and chicken are not significant relative to
bakery. Overall, what do we learn from this output? We know now that several categories, such
as fountain sodas, bakery items, breakfast sandwiches, help increase the chance that a
customer will be a loyalty customer. Tikka could promote these products to try to draw in more
loyalty customers.

116
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-3.4 Logistic Regression Hands on - One Variable (Part3)

So our next step is to evaluate the model that we've created by predicting the probabilities of
each row and examining how accurate the model is at predicting a loyalty purchase. Recall that
logistic regression uses the independent variables to create a probability for each row that
maximizes the likelihood that the row is close to the true category, loyal or not loyal. Whichever
one is the true category. We should predict those probabilities and examine their accuracy since
we've labeled data and we know what the true category is for each row. So first. Let's go ahead
and predict the probabilities. We really want to do this with the testing data. But we'll do this with
the training data first and then compare to the real data and the testing data after that.

117
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So, let's examine this code here, line by line. The first line is the predictions of the probabilities.
We're predicting using the model we just created, model train. All right. So we have the
predictive function. We're going to call this new prediction object predict train. Again, we have
the training data. Next, as you can see, we make sure we're using the training data by having
new data equal data train. The next argument, type equals response, gives the predicted
probabilities rather than the law gods, which are more difficult to interpret. So next we're going
to print a summary of the probability just so that we can get a sense of what they look like. To
further evaluate the model, we take the probabilities, predict underscore train and we're going to
put them into the training data as a new column called prediction. Finally, we're going to take a
look at a snapshot of the data. So let's run this. We train up our model, we go ahead and do
head to look at the top ten rows of our training data. All of these first ten rows are from the
loyalty customers as we look. They're all from loyalty and they're all relatively high under the
prediction column here. Considering that all of them are are above 58% or so.

118
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So next we want to evaluate our model in another way by looking at how accurate the model is
in making correct predictions. So there's a lot of ways to think about accuracy of the model. We
can think about how accurate the model is at detecting loyalty, but you can also think about how
accurate the model is at detecting not loyal or a non loyalty transaction. Both of these types of
accuracy makes sense since we want the overall accuracy of the model to be really good. A
typical way to visualize accuracy is by using a confusion matrix or a classification matrix. Again,
we're using the training data here. So just so we can compare later on to the testing data which
we want to use ultimately. So let's again look at the code line by line.

119
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The first line creates a table using the table function that we saw before. All right. This table
takes the probability predictions from above, predict underscore train and writes false if the
probability is less than 0.5 and true if it's more than 0.5. These are our predictions using again
the training data. So let's run that and you can see here at the top what we have. So you can
see the faults and true categories here on the left of the table. So it's going to compare these
with the truth, which is the loyalty column. So not loyal and loyal here on the top. So the second
line then up here from the code is going to run the function from way above that we used at the
beginning and ran at the beginning of the notebook.

120
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

And this function just takes the four cells of this table and labels them and manipulates them to
create various types of accuracy displayed right here below. So you have that all here and let's
go ahead and explore this. The table shows the following output. First, when a transaction is
actually truthful, a loyalty transaction, the model correctly classifies it

as such by saying true. So that's right. That's this right here. The 185, 185,131.

121
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So this is called a true positive or a hit. Now, when a transaction is truthfully a non loyalty
transaction, okay,

so the truth is not loyal and indeed our model says fault says It's not loyal. That happens,
265,829 times. This is called a true negative or rejection and again, that's also listed right down
here.

122
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

On the other hand, the model makes two kinds of errors. When the model transaction is not a
loyal transaction, but the model incorrectly says that it is, this is called a fault positive or a type
one error. And this happens 124,171 time. So that's here.

Again, this is not a loyalty transaction because this is truth. But the model says that it is and so
that we call that a type one error.

123
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Finally, when a transaction is a loyalty transaction and the model says it is not, which here
happens 204,677 times it's called a faults negative or a type one error.

Again, we see that here where it actually is a loyalty transaction but the model says it is not.

124
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So these numbers can then be manipulated, multiplied and divided to create different measures
of accuracy that are commonly used in practice. For example,

overall accuracy is probably the most important thing that we're interested in here. Overall
accuracy is 57, about 58%.

125
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next we have sensitivity. This is also called the recall rates or the true positive rates or the hit
rate. Basically, how many positives did the model get right.

And here we're actually below 50% at 47%. Next we have specificity also called selectivity, the
true negative.

126
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

This basically answers how many negatives did the model get right. And here we're at 68%.

Next we have another common measure of accuracy and that's precision. Also called the
positive predictive value. This basically says how good are the model's positive predictions. And
again, in each case I've indicated how these are calculated, so you can refer back to that. Last,
we have the negative predictive value, and this here is basically saying how good are the
model's negative predictions.

127
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Thus, overall, our model predicts better than chance, that overall accuracy level was about 57,
58%. Recall that loyalty transactions happen about 50% of the time. Thus, if we had no model at
all [LAUGH] and just flipped the coin every time to guess whether a transaction was going to be
a loyalty transaction or not, we'd get 50%. We'd get it right about half the time. So our model
has helped us to get it right a little bit more than that. It helps us get it right 58% of the time,
which is not super high, but is better than chance. Where the model really kind of falls down and
is not great, is in the sensitivity area. That is, the model is pretty good at predicting not loyal,

128
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

68% right here. The specificity is good but bad at predicting loyal, getting it right, only 47% of
the time. That's worse than flipping a coin. So our next step is to improve this model.

However, before we do that, let's take our training model and predict or run it on the testing
data. Recall that we train on the training data, we test on the testing data. So we want to use
new data to test on so that we're not just trying to memorize the training data, so that we don't
over fit our model. But rather we're creating a model that we can work on any data.

129
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So let's do that now and run this. This code of course, looks very similar. So what we did above,
the key exception here is that the data being used is now the test data. So that's the big critical
key difference that you don't want to miss. So how does our model perform in the new data.
Well, it actually does a little bit better than the training data. So while accuracy is is pretty similar
here, you know, still about 58%. Our problem from before, the low sensitivity is slightly
improved. Not much, but slightly improved.

130
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon
Lesson 2-3.5 Logistic Regression Hands on - Multiple Variables

In this video, we'll continue exploring logistic regression. Logistic regression is a valuable
classification type machine learning algorithm. In this video, we'll use the algorithm to predict
whether a transaction is from a loyalty customer or not. We'll use the TEKA data set. Now,
TEKA desires to increase the loyalty customers that they have and the number of products that
their loyalty customers purchased because loyalty customers creates more business and more
profitable business for them. As always, we encourage you to resist the urge to just watch the
videos rather, please follow along, doing each of the steps that we're doing so that you can get
the critical hands-on experience necessary to learn this valuable tool. First, let's bring in the
data and the needed packages. Last time we predicted loyalty using a single variable category
but this time, we'll use all of our variables. We'll see if this improves the prediction accuracy of
our model. Before we get to the model statement, the code will be identical to the prior video
when we used only the one variable so we'll run all of that here just right now

131
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

all the way up to the model statement. Of our six steps for using a machine learning algorithm,

we're now focused on step 5. The first step is data scrubbing, the second was algorithms
selection, third was model training or fitting, fourth was model evaluation, and five is model
improvement, six is deploying the model, so we're firmly into the number five here.

132
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

All right. So as you can see as you scroll through this code again, this is just identical from last
time when we had a univariate logistical regression model. Now, with a multivariate regression
model with multiple independent variables, everything will be the same until we get to the model
statement.

Great. So our data split. We're now ready to train our model. Recall that an algorithm is a set of
rules for how to make a model. A model takes our data and options that we provide it to select
and create new data and information and additional rules for how to use the model.

133
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Below, we create our model called model_train and then summarize it to view the output. So
we're again, using the GLM function with family equal to binomial to indicate that it's a logistic
regression. Our dependent variable continues here to be loyalty, but this time will include all of
our independent variables in the model, not just category. This is multivariate logistic regression
so we have multiple independent variables. We are, of course, using the training data here,
since we're training the model. So let's go ahead and run this and summarize as we train our
model using both quarter, category and state. All right. So here's our output. Now, last time we
examined the impact of category on loyalty. So let's look this time at quarter, right down here.

134
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

The quarter is just a quarter of the year. We made it into a factor here so that R knows the
category. There's some debate about whether we could leave this as a continuous variable or
not, since the difference between one quarter of the year and another quarter of the year does
have meaning but for our purposes, we'll leave it as a factor or as a categorical variable. The
first quarter of the year is the reference level of these multiple dummy variables. Thus, when we
evaluate the effects of these quarters, it's in relationship to the effect of quarter 1. Looking at the
signs of the coefficients and the p-values that are highlighted here, we can see that each
quarter significantly increases the probability that a transaction will be a loyalty transaction. For
example, the third quarter of the year has a positive coefficient

135
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

of 216750 and a very low p-value less than two with a 16 zeros in front of it.

This means that relative to, again, quarter 1, quarter 3 has more transactions that are from the
loyalty customers. With this knowledge, TEKA could consider focusing on the beginning of the
year in an effort to increase the number of loyalty customers that they have. Additionally, TEKA
could focus on the summer and fall as key times when their loyalty customers all are already
active to promote loyalty to them and help increase those loyalty customers' purchases even
more. Next, let's look at how each state affects the occurrence

136
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

of our loyalty transactions. These are the dummy variables here. The reference state, the state
up and intercept, is Alabama.

So relative to Alabama, the states that are most likely to generate loyalty transactions are
Colorado, Missouri, and Wyoming. Missouri, in particular, has a strong effect. This can be seen
by their relatively large coefficient. So TEKA should investigate key stores in this state to see
what's going on right here. Finally, let's examine our larger, more complicated model on our test
data. Recall that our single variable model had a relatively modest accuracy rate of about 57, 58
percent. The more complex model, let's look at that now.

137
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

So again, here we've included our test data because we're going to run this model on the test
data. We want to see how well our trained data predicts on our test data. We've run summary
right here so we can view that, and then we've added our predict tests, our probabilities, back to
our data set. If we wanted to, we could look at that and we see it right here. Then we've run
heads so we can look at these. Finally, we are going to run this confusion matrix that we looked
at last time. The more complex model has increased the accuracy slightly up to about almost 60
percent as you can see here but perhaps more importantly, the sensitivity of the model is much
improved right here. Now, recall that before univariant logistic regression we had, the sensitivity
or the accuracy of that mode predicting not loyal was below 50 percent,

138
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

worse than a coin flip, but this new model has increased that all the way up to 57 percent. So
overall, this is not a great model. We could certainly add additional variables to make it more
complex. Additionally, recall that this data is just a subset of the available data so adding data
would also improve the predictability of our model.

139
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Module 2 Review

Module 2 Conclusion

Congratulations. In this module we've really covered a lot of ground. First, we've done an
overview of machine learning to provide you with a solid framework or foundation for how to
understand machine learning. Thus in the future, in this class or your own learning efforts, when
you encounter a new model, that you've never encountered before, you'll be able to slide into
the framework, understand it quickly, and move more quickly through the process of
understanding the new algorithm. More importantly, you'll be able to put that algorithm to work
for you in solving business problems. In this module, we did the following. We discussed
different uses for machine learning, inference, prediction, and experimentation. In all three of
these paradigms, data is king. Data and our statistical analysis of data drives our decisions.
Next, we examine the building blocks for machine learning models, data and data types. We
explored the differences between categorical, and continuous variables.

140
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Next, we talked about the different types of machine learning, and machine learning algorithms.
We talked about supervised, unsupervised and reinforcement learning, their definitions and their
differences. After that, we discussed regression, classification, and clustering algorithms and
how they differ. We then showed a graphic to put everything in its place.

We then walked through a series of six steps for machine learning. We also discussed the
difference between machine learning algorithms, and models.

141
 Machine Learning Algorithms with R in Business Analytics
Professor Jessen Hobson & Ronald Guymon

Finally, we introduced a business problem from our hypothetical company TECA. We learned a
classification algorithm, logistic regression and used the algorithm, and the model to gain insight
into when TECA's transactions are made to loyalty versus non loyalty customers. This can help
the company increase their number of loyalty customers, as well as the number of transactions
their loyalty customers make.

142