MACHINE LEARNING FOR

BEGINNERS
See more books: https://www.auvapress.com/books

Leave a review on Amazon:

https://www.auvapress.com/amazon-review/machine-learning-beginners
MATT GATES
© Copyright 2017 Matt Gates - All rights reserved.
In no way is it legal to reproduce, duplicate, or transmit any part of this
document in either electronic means or in printed format. Recording of this
publication is strictly prohibited and any storage of this document is not
allowed unless with written permission from the publisher. All rights
reserved.
The information provided herein is stated to be truthful and consistent, in
that any liability, in terms of inattention or otherwise, by any usage or abuse
of any policies, processes, or directions contained within is the solitary and
utter responsibility of the recipient reader. Under no circumstances will any
legal responsibility or blame be held against the publisher for any
reparation, damages, or monetary loss due to the information herein, either
directly or indirectly.
Respective authors own all copyrights not held by the publisher.
Legal Notice:
This book is copyright protected. This is only for personal use. You cannot
amend, distribute, sell, use, quote or paraphrase any part or the content
within this book without the consent of the author or copyright owner.
Legal action will be pursued if this is breached.
Disclaimer Notice:
Please note the information contained within this document is for
educational and entertainment purposes only. Every attempt has been made
to provide accurate, up to date and reliable complete information. No
warranties of any kind are expressed or implied. Readers acknowledge that
the author is not engaging in the rendering of legal, financial, medical or
professional advice.
By reading this document, the reader agrees that under no circumstances are
we responsible for any losses, direct or indirect, which are incurred as a
result of the use of information contained within this document, including,
but not limited to, —errors, omissions, or inaccuracies.
Trademarks:
Auva Press and the Auva Press logo are trademarks or registered
trademarks and may not be used without written permission. All other
trademarks are the property of their respective owners and use in an
editorial fashion to the benefit of the trademark owner, with no intention of
infringement of the trademark. Auva Press is not associated with any
product or vendor mentioned in this book.
FIRST EDITION
ISBN-13: 978-1-5470-3904-3
ISBN-10: 1-5470-3904-3

Editor: Audrey Swans

Cover Designer: Howard West
 To my dearest and family,
your support have been the great source of strength
and making this possible
CONTENTS

PREFACE
Chapter 1 Introduction To Machine Learning
What is learning?
When is machine learning important?
Applications of machine learning
Chapter 2 Introduction to Statistics and Probability Theory
Random variables
Distributions
Mean, variance, and standard deviation in statistical distribution
Marginalization
Bayes’ theorem
Chapter 3 Building Blocks of Machine Learning
Formal statistical learning frameworks
Empirical risk
PAC learning strategies
Generalization models for machine learning
Chapter 4 Basic Machine-Learning Algorithms
Challenges of machine learning
Types of learning
Chapter 5 Supervised Machine-Learning Algorithms
Decision trees
Random forest
KNN algorithm
Regression algorithms
Chapter 6 Unsupervised Machine Learning Algorithms
 Clustering algorithm
Markov algorithm
Neural networks
Chapter 7 Reinforcement Machine Learning Algorithms
Q-learning
SARSA
CONCLUSION
FURTHER RESOURCES
ABOUT THE AUTHOR
PREFACE

Machine learning has become the pillar of information technology (IT),

albeit hidden over the past few years. Today, we are surrounded by several
machine learning based technologies across a broad spectrum of
applications, such as credit card fraud detection systems, software systems
that filter out spam messages, and search engines that are always learning to
bring us better results.

Unlike conventional programs, machine-learning programs utilize human

like artificial intelligence (AI) to learn from experience rather than just
follow explicit instructions, which enables these programs to adapt on their
own and detect more complex data patterns during the computational
process. This is not possible through conventional programing techniques.
As ever-expanding volumes of data are collected, “smart” data analysis will
become ubiquitous and an integral component of any technological system.

In the following chapters, you will learn all the ins and outs of machine
learning over a wide range of applications, exploring some of the basic
tools from statistics and probability theory that define how machine-
language problems are framed and understand fairly basic yet effective
algorithms that solve complex machine-learning problems. Let’s jump in.
 Chapter 1

Introduction To Machine Learning

In machine learning, also known as automated learning, computers are

programed to learn from any input that they are given and provide answers
to complex problems. The input for any learning process or algorithm is the
training data, which represents the experience, while the output is some
complex task performed by a software program.

But more specifically, the biggest query that should concern us at this stage
is “How do we represent the training data or input so that experience can be
described?”

And how can the process of automated learning be achieved in programs?

In this chapter, we delve deeper into these concepts of machine learning to
give you a general idea of machine learning.

Let’s dive in.

What is learning?
Learning can be described as any process that converts experience into
knowledge or expertise. Let me explain using a couple of examples.

For example, when mice come across an unfamiliar food that looks or
smells good, they eat it in small quantities at first. If the food tastes bad or
makes them sick, the mice associate that food with illness and consequently
avoid it from then on. Likewise, this behavior can be found in other animals
in the wild when venturing into new territories.

On the other hand, if the food doesn’t have any ill effects on their bodies,
the mice continue eating the food. Obviously, a learning process has taken
place: the mice have used their experience eating the food to acquire basic
knowledge of the safety of that food, automatically predicting that it will
have the same effect on their bodies when encountered in the future.

Another example of learning is training dogs to do an action like rolling on

the ground or simply sit and await instruction. Every action or task
completed successfully is rewarded with food. This reinforces the learning
process of the dog through repetition training. The dog learned that food
will rewarded for task been done correctly upon instruction and will follow
this instruction even if there are no presence of food.

Let us now examine a typical machine-learning process. Suppose you

would like to program a computer to learn how to filter out spam e-mails. A
simple solution would resemble the learning process that mice use to avoid
poisonous foods: you would just program the computer to identify,
categorize and memorize all the e-mails that the user labeled as spam, then
search for identical e-mails among the new ones. If a new e-mail matched
one on the spam list, it would automatically be labeled as such and trashed.
Otherwise, it would be stored in the user’s inbox.

While the process of learning by memorization is sometimes worthwhile, it

lacks an integral component of any efficient learning process—inductive
reasoning, the ability to learn broad generalizations from specific examples.
In other words, it would be better to program the computer to learn how to
discern the types of e-mail messages that the user considers to be spam. To
achieve such generalization in the spam-filtering process, the program
could scan the known spam e-mails and mine a set of words that often
appear in spam messages. By checking new e-mails for these suspicious
words, the program could then correctly predict whether an unfamiliar e-
mail message would be labelled as spam.

However, inductive reasoning shouldn’t be the only approach that machine-

learning systems use.

What will distinguish learning mechanisms in spam filtering program that

just result in new emails being labeled as spam or stored in the box from
useful learning process?

Answering this question is vital to the development of machine learning

systems. While people can depend on common sense to filter out the
random spam messages, computers don’t have common sense.

Therefore, so we must define refined principles that protect the programs

need refined principles to prevent them from reaching illogical and
meaningless conclusions. The development of these principles is what
underpinnings of machine learning program design.

To summarize, prior experience combined with principles biasing the

learning task is the key to creating a successful learning process. Therefore,
every well-designed machine-learning system requires tools that express a
domain expertise, translate this expertise into a learning bias, and quantify
the bias’s effect on the success of learning.
When is machine learning important?
Under what circumstances do you need a machine-learning system rather
than a conventional program that ordinary computers can handle? There are
two types of problems that machine learning can solve:

Tasks that are too complex to be programed

Need for adaptively generated tasks

Let’s explore these circumstances.

1. Tasks that are too complex to be programed

For various reasons, not all tasks can be described using conventional
programing languages; some tasks are too complex in their requirements or
lack clarity in defining the application interface. We can group these tasks
into two sets:

Tasks that are routinely performed by people and animals. There

are several tasks, such as driving, speech recognition, and image
recognition, that we perform regularly in everyday life and yet
cannot describe sufficiently enough to create a conventional
program. In all these tasks, programs have to learn from experience
to achieve satisfactory output, which requires machine-learning
systems.
Tasks that are beyond human capabilities. These tasks involve
complex analysis to turn large sets of data into useful information
for decision-making. For instance, medical-archive data, weather-
prediction data, genomic data, and search engines are all too large
and unstructured for conventional SQL to process. But as more and
more data are digitally recorded, it is becoming increasingly evident
that a treasure-trove of meaningful information is buried in these
 databases, and the only way to find it is by developing machine-
learning systems to help mine the data. If you would like to know
more about data analytics and mining, the previous book in this
series, Data Analytics for Beginners, discusses these topics in
greater detail.

2. Need for adaptively generated tasks

One of the limitations of conventional programs is their rigidity; once these
programs are coded, tested, and installed, they remain unchanged,
performing the same tasks over and over again. But numerous tasks vary
from user to user, and machine-learning programs are better equipped to
perform these tasks, since they can adapt their behavior to different contexts
and environments. Typical examples of such programs include:

Programs that decode handwritten text, easily adapting to the

various handwriting styles of different users
Spam-detection programs that adapt automatically to any changes
in spam e-mails
Speech-recognition programs that readily adapt to the various
speech patterns of different users
Applications of machine learning
Machine learning is a fantastic revolution in AI. Today, both cutting-edge
startups and established firms are finding more and more ways to utilize
machine learning in their business operations. Here are some examples of
machine-learning applications to various fields:

#1: Finance
In the financial world, online learning and decision-making take place in a
wide variety of scenarios that can change depending on the business
environment, the available feedback, and the nature of the decision. These
decisions can involve stock trading, ad placement, route planning, or even
picking a heuristic move.

The machine-learning algorithms that are used in online stock trading

ensure the best possible outcomes in two ways. First, they help prevent
stockholders from making the worst trades. Second, their benchmark
principles narrow down decisions to the most profitable strategies.

Machine learning can also help banks discover important new insights in
data, enabling them to compete better and increase their bottom lines. Such
data mining assists not only in determining the best investment
opportunities but also in identifying high-risk customer profiles and
detecting signs of fraud.

#2: Big Data and Analysis

Today, data scientists are having a lot of problems efficiently storing the
vast amounts of data in their organizations, let alone learning anything from
them. Data collection has been growing exponentially at a rate of over 45
percent every year, and an astonishingly large portion of this growth is
stemming from either unstructured or semi-structured data that can’t be
handled by the conventional relational database-management systems.

Managing such data requires a disruptive approach, and machine learning

hasn’t disappointed in this regard. From medical fields to retail stores to
election-campaign management, machine learning is helping organizations
derive new insights from data, which they then use to tailor their services to
their clients’ behavioral patterns.

For instance, online storefronts are now using machine learning to

recommend items that you may like based on analysis of your purchasing
history. Similarly, machine learning has enabled the healthcare industry to
develop wearable devices and sensors that use large volumes of data to
assess patients’ health in real time, reducing readmission rates at hospitals
and helping healthcare professionals improve their diagnoses.

#3: Statistical research

Machine Learning has become the pillar of IT, albeit hidden over the past
few years. From search engines that are constantly learning to bring us the
best results to anti-spam software systems that filter out spam messages to
credit card fraud detection systems, the list of machine learning applications
is endless.

One universal trait of these applications is the complexity of data patterns

that should be analyzed during the computational process. The conventional
programming techniques are inadequate because they don’t provide explicit
and refined user requirements that specify how the tasks should be
computed. In particular, machine-learning applications should utilize
intelligence such as that possessed by human beings to learn from
experience rather than just following explicit instructions.
Such intelligence is a necessary and integral component that should help the
conventional programs with capabilities of learning and adapt on their own.
It is no secret that machine learning will continue to expand and influence
virtually all spheres of our lives. With the ever-expanding volumes of data,
there’s a good reason why data analysis smart will become ubiquitous and
an integral component of any technological system.
 Chapter 2

Introduction to Statistics and Probability

Theory

You may be wondering, “What is the relationship between probability

theory and machine learning?”

Well, Machine learning is a broad field that intersects several other fields
such as statistics and probability theory, AI and computer science
algorithmic field that helps to build programs that can iteratively learn on
their own and extract meaningful insights from large sets of data. To unlock
the immense potential of machine learning in a given application, a
thorough understanding of statistics and probability theory is necessary for
a good grasp of machine learning to get the best intelligent systems.

Obviously, there are several reasons why statistical and probability theory is
essential ingredients of machine learning. Some of these reasons are:
 Selecting the algorithm that provides the best balance of accuracy,
number of parameters, training time, and model complexity;
Selecting the appropriate parameter settings for the program and its
validation strategies;
Identifying the bias-variance tradeoffs that cause underfitting and
over fitting; and
Estimating the right confidence interval and level of uncertainty in
the program’s decisions.

Therefore, this chapter explores concepts of statistics and probability theory

that are essential in machine learning.
Random variables
Random variables play a significant role in probability theory. In
probability theory, a random variable, usually denoted as X, is a variable
whose possible values are primarily numerical outcomes of a random
experiment or phenomenon. Random variables can be either discrete or
continuous, which means they are actually functions that can map outcomes
to real values within the outcome space. Let me explain using the following
example:

Suppose you throw a die. Let X represents the random variable that relies
on the outcome of the throw. Obviously, the natural choice for X would be
to map the outcome denoted as i to the value of i. An X outcome of 1, for
instance, would map the event of throwing a one on the die to the value of
1. (We could also choose strange mappings such as a variable Y that maps
all outcomes to 0, but that would be a tedious and boring function.) The
probability (P) of outcome i of random variable X is denoted as either P(X
= i) or PX(i); in this way, you can avoid the formal notation of event spaces
by defining random variables that capture the appropriate events.
Distributions
Probability distribution is the probability of each outcome of a random
variable. Consider the following example:

Let random variable X again represent the outcome space of the die throw.
Assuming that the die isn’t loaded, we would expect the probability
distribution of X to be as follows:

PX(1) = PX(2) = ... = PX(6) = 1/6

While this example is similar to the previous one, it has a different

meaning: probability distribution is defined by the spectrum of events,
whereas the previous example is defined by the random variable.

In probability theory, P(X) usually denotes the probability distribution (P)

of random variable X. A distribution can include multiple variables
simultaneously, in which case it is called a joint distribution because the
probability is determined jointly by all the variables that are involved.

For example, let X be a random variable defined as the outcome space of a

die throw and Y be a random variable that indicates the outcome of a coin
flip, assigning a value of 1 if the coin turns up heads and a value of 0 if it
turns up tails. Assuming that both the coin and the die are fair, the joint
probability distribution of X and Y would be as follows:

P X=1 X=2 X=3 X=4 X=5 X=6

Y=0 1/12 1/12 1/12 1/12 1/12 1/12
Y=1 1/12 1/12 1/12 1/12 1/12 1/12
We would denote this joint distribution as P(X,Y) and the probability of X
having an outcome of a and Y having an outcome of b as either P(X = a,Y =
b) or PX,Y(a,b).

Conditional distributions
Using a conditional distribution, the distribution of a random variable when
the value of another random variable is known, we can base the probability
of an event on the given outcome of another event. The conditional
probability of random variable X = a given that random variable Y = b is
defined as follows:

P(X = a|Y = b) = P(X = a,Y = b)/P(Y = b)

For example, suppose we want to determine the probability of throwing a

one on a fair die given that the number thrown is odd. Let X be a random
variable indicating the number thrown on the die and Y be a random
variable that assigns the value of 1 to the event of throwing an odd number.
The probability of throwing a one can then be calculated as follows:

P(X = 1|Y = 1) = P(X = 1,Y = 1)/P(Y = 1) = (1/6)/(1/2) = 1/3

Conditional distributions are an important tool for designing systems that

can reason about uncertainty.

Independence
A random variable is independent of another random variable when the
variable’s distribution doesn’t change in response to the value of the other
variable. In machine learning, we can make assumptions about data based
on independence. For instance, training samples i and j are assumed to be
independent of an underlying space when the label of sample i is unaffected
by the features of sample j. In formal notation, random variable X is
independent of random variable Y only when
 P(X) = P(X|Y)

The values of X and Y have been dropped because the statement is true for
all values of X and Y.
Mean, variance, and standard deviation in statistical
distribution

Mean
Also known as the expected value or the first moment, the mean (E) of the
distribution for each outcome (a) of random variable X is defined as
follows:

E(X) = ∑ [aP(X = a)]

For all values of a being, members of values of variable X.

In other words, the mean of distribution can be given by the formula below:

Where µ is the mean of the distribution. Consider the example below:

For example, let X be a random variable representing the outcome of

throwing a fair die. The expected value of X can then be calculated as
follows:

E(X) = [(1(1)1/6)] 6 + [(2(1)1/6)] 6 + ...· · · + [(6)(1 1 /6)]6 = 3.5 1/ 2

The mean of a random variable X is useful for assessing the expected of the
variable’s distribution. For instance, if you’re a stockbroker, you may want
to know the expected value of a client’s investments in a year’s time. You
may be want to investigate the risk of your investment. In other words, how
likely it is that the value of your investment will deviate from its
expectation?

Variance
To protect the client from unacceptable financial losses, the stockbroker
may also want to estimate the amount of risk involved in an investment by
calculating its variance. The variance or second moment of a probability
distribution is the degree of disparity within the distribution, defined in
formal notation as

Var(X) = E {[X − E(X)]2}

where X is a random variable and E is the mean of the distribution.

However, variance is much easier to compute using the following formula:

Var(X) = E(X2) – [E(X)]2

Instead of

Var(X) = E ((X − E(X)) 2)

Standard deviation
The variance of a random variable’s distribution measures only the average
degree of disparity, which is why its use is mostly confined to calculating
risk. To more accurately determine how much the random variable deviates
from its expected value, the standard deviation is used instead. The standard
deviation (σ) of the distribution of random variable X is the square root of
the distribution variance:

Where is the standard deviation of the probability distribution.

Marginalization
Suppose the random variables X and Y comprise a joint distribution,
P(X,Y), and we want to isolate the distribution of X, P(X); we can
accomplish this either by summing out Y if X is a discrete random variable
or by integrating Y if X is a continuous random variable. Such an operation
is called marginalization.

P(x; y)

This means that given the joint density, we can recover the P(x) by
summing it out with y if it is a discrete random variable or integrating it
with y if it is a continuous random variable. Such an operation is called
marginalization.

Put simply, if X is a discrete random variable, then we can have the

following distribution:

p(x) = ∑ p(x; y)

For all values of y. In such a scenario, we say that the random variables X
and Y are independent which mean that the values that X assumes don't
depend on the values that Y assumes.

p(x; y) = p(x)p(y)

Independence is vital when you’re dealing with large numbers of random

variables whose behavior you want to estimate jointly. For example,
whenever you do repeated measurements of a quantity such as measuring
the voltages of a device, you can assume individual measurements and
estimate the large sets of devices that you’re dealing with.
On the other hand, dependence can be useful for classification and
regression problems. For example, the traffic lights are dependent on each
other at an intersection. If you’re a driver, you can infer that when the lights
are green in your direction, there won’t be traffic crossing your path since
the other lights will be red. Equally, whenever you are given a picture x of
any digit, you hope that there will be some dependence between that picture
and label y.

This can formally be represented as:

p(x|y):= p(x; y)/( p(y))

Bayes’ theorem
Using the axioms of conditional probability, Bayes’ theorem describes the
process of updating the probabilities of hypotheses when new evidence is
uncovered, but it can also be a powerful tool for reasoning about a broad
range of problems that involve belief updates. Given two events, E and H,
Bayes’ theorem is defined in formal notation as follows:

"P" ("H" │"E" )" = " ("P(H)∙P" ("E" │"H" ))/"P(E)"

The applications of Bayes’ theorem have produced a new field of

probability theory called Bayesian statistics, revolutionizing scientific
research but also medicine, law, quality analysis, and every other discipline
that draws inferences from uncertain evidence.
 Chapter 3

Building Blocks of Machine Learning

Machine-learning algorithms have attracted more interest than ever before,

generating a flood of innovative ideas for applications of this technology.
For tech enthusiasts interested in developing machine-learning programs,
the greatest challenge is about understanding the fundamental concepts
involved in these projects.

In this chapter, we explore the building blocks of machine learning that will
help you close the gap. By the end of the chapter, you should be in a
position to understand the finite elements of machine learning that can help
you turn ideas into hardware and software primitives.
Formal statistical learning frameworks
Statistical learning contexts can simplify learning processes in computers.
Allow me to illustrate using an example.

Suppose you’ve landed on an island where the natives love to eat papaya, a
fruit with which you are unfamiliar. You’re standing in the marketplace,
trying to figure out which papayas are the best. Based on your previous
experiences with fresh fruit, you decide to examine the color and softness of
each papaya. The colors range from red to dark brown, while the softness
ranges from rock solid to mushy; in statistical terms, these are two sets of
inputs to help you predict the taste of each papaya.

Let us now construct a model to describe this learning process, which is

called a formal statistical learning framework. The model must incorporate
the following components:

Learner’s input
Learner’s output
Simple data-generalization model
Measures of success

Let’s examine each of these critical elements.

#1: Learner’s input

The domain set. The domain is an arbitrary set containing the
objects, also called points, to be labeled. In this case, the domain set
is the set of all papayas. Domain points can be represented using
vectors of features such as a papaya’s color and softness.
The label set. The label set contains the predictions made about the
domain points. For this scenario, the label set can be restricted to a
 two-element set where Y represents the papayas you think will taste
good and X represents the papayas you think will taste bad.
The training data. The training data is a set containing ((x1; y1):::
(xm; ym)) that has a finite sequence of pairs in the X×Y matrix that
are accessed to test the accuracy of the label set. In this example,
the training data is the set of papayas you buy so you can find out
how they taste.

#2: Learner’s output

The output is the creation of the prediction rule, also called the predictor,
hypothesis, or classifier, used to label the domain points. In this example,
the prediction rule is the standard you use to decide whether you think a
papaya will taste good or bad. Prediction rules are typically based on
previous experiences.

#3: Data-generalization model

The data-generalization model uses a function to generate the training data
based on the probability distribution of the domain set. In this case, the
data-generalization model is the method by which you choose a
representative sample of papayas to buy. However, the learner might not
always know what the distribution is, so the model’s design must ensure
that the labels apply to any probability distribution. For instance, the
function can take either of the following forms:

f:X!Y

or

Yi yi= f(Xixi)

for all values of i.

Such a function can now be used to generate the training data that the
learner will now use.

#4: Measures of success

Obviously, there must be a way to calculate the training error or error of the
predictor, the probability that the statistical learning framework incorrectly
predicts the outcomes of the learning process and thus fails to meet the
desired objectives. Since the objective of this scenario is to select the best
papayas, the error of the predictor can be determined by conducting a taste
test of the papayas you bought to find out whether they’re as good or bad as
you expected them to be.
Empirical risk
In the situation sometimes called agnostic learning, a machine-learning
algorithm receives input from a training set that is sampled from an
unknown distribution. The goal of the machine-learning algorithm is to
minimize error, but since the learner doesn’t know what the distribution is,
the exact computational error cannot be determined.

Any useful notion of the error that can be computed by the learner will be
the training error, which is the error that the classifier gets from the training
sample.

In summary, the minimization risk R(h) can’t be computed since the

distribution D is unidentified to the machine- learning program under such
circumstances—a situation that is sometimes called agnostic learning—and
the function is minimized. However, we can use an approximation which is
the empirical risk. Every learning model has a loss function that measures
the degree of inaccuracy in the model’s predictions. To calculate the
empirical risk, the smallest values in the training set are plugged into the
loss function, which is called minimizing or leveraging the function.
PAC learning strategies
In the previous sections, we have seen how a finite hypothesis class and the
training data set rule can be applied to train a computer on how to adapt to
certain tasks whose size is independent on the underlying distribution D and
the unknown function f. However, we haven’t defined the role of “Probably
Approximately Correct” (PAC) learning.

At the core of probably approximately correct (PAC) learning are two

approximation parameters: First, the accuracy parameter and the second is
the output classifier.

The accuracy parameter determines how often the output classifier

can make a correct prediction.
The confidence parameter measures the probability that the
predictor will attain this level of accuracy.

Although PAC learning can be useful even when a model’s training data
accurately represent the distribution, it is especially vital for handling the
inevitable uncertainties of the data-access model. This model always
generates random training sets with a small chance of producing an
uninformative training set that has only one domain point.
Generalization models for machine learning
The data-access model that we have discussed in the previous section can
readily be use to represent the wider scope of learning processes in general.
By generalization, we mean having two essential components:

Removing the reliability assumption

Revisiting the true error rate

When the reliability assumption has been met, we can expect a machine-
learning algorithm to generate training data that reliably reflect the
distribution. However, such an assumption can be impractical, placing
unrealistic standards on the algorithm.

Moreover, a learning algorithm is never guaranteed to find a hypothesis

with the best possible error rate like that of the Bayes predictor. Once we
make no previous assumptions about data-generating distribution, there is
no algorithm will guarantee the production of predictors that are as good as
the Bayes optimal one.

Instead, it is reasonable to expect the algorithm to find a predictor whose

error rate isn’t much less than optimal.
 Chapter 4

Basic Machine-Learning Algorithms

This chapter explores some problems commonly encountered by machine-

learning programs in various applications and introduces the types of
algorithms that are used to solve these problems.
Challenges of machine learning
The predominant theme of all machine-learning problems is finding
relationships between sets of data. However, the specific problem that
machine learning seeks to solve depends on the application, so the
following synopses briefly describe typical program designs for some
common applications and the challenges inherent in these designs:

1. Search engines.
Probably the most familiar application of machine learning, search
engines are designed to find the Web pages that are most relevant to a
user’s query. To accomplish this goal, a search engine needs to figure
out which pages are relevant and which ones exactly match the query.
The program can draw its conclusions from a variety of inputs, such
as the words in each link, the content of each page, or how frequently
users follow the suggested links.

2. Collaborative filtering.
Collaborative filtering is the learning process by which online
retailers such as Amazon recommend items that customers may be
interested in buying. These programs are similar to search engines,
but relevancy is based on trends in a user’s past purchases rather than
on direct input from the user.

3. Automatic translation.
A particularly challenging task in machine learning is automatic
translation of documents. It’s one thing for a computational linguist to
develop a curated set of rules for a program to follow, but it’s quite
another for the program to understand the words in a document and
learn other rules of syntax and grammar on its own, especially since
documents aren’t always grammatically correct. One solution is to
design a program that learns how to translate languages by using
examples of translated documents.

4. Named identity recognition.

The challenge of identifying entities such as places, titles, actions and
names from documents can easily be modeled and developed in
machine learning systems. These tasks are vital in the automatic
digestion and comprehension of documents. For instance, modern
email clients such as the Google’s Gmail and Apple’s Mail are
nowadays being shipped with the ability to identify email addresses in
them and filling them in address the phone’s address book.

While machine-learning systems that use handcrafted rules can

produce satisfactory results, it is far more effective to use examples of
other marked-up documents to learn from such dependencies
automatically so that the machine learning system is efficient in
named identity recognition.
5. Speech recognition
The predominant theme of all machine-learning problems is that there
exists a nontrivial dependence between some observations that we
commonly refer to as x and the desired response that we commonly
refer to as y for which set of rules are not known. By using machine
learning, we can infer some of these dependencies between x and y
using algorithms that can associate between these two sets of data.

For instance, we can annotate any audio sequence with the text data
such as Apple’s Siri, the recognition of handwriting and signatures
and the avatar behavior that is used in computer games. This
information had increasingly led to the development of robotic
systems that are far much more efficient in work than when these
systems were not there.

6. Facial recognition.
These days, facial recognition is a primary component of many
security and access-control systems. Using photos or video
recordings, such a system is designed to allow only people it
recognizes into the building or information system it’s guarding. The
difficult part, of course, is verifying a person’s identity; a machine-
learning program is far more effective in this regard, since a system
that learns from its mistakes is much harder to fool than one with a
rigid set of standards.
Types of learning
Machine learning is a broad field with several different subfields of learning
processes. These subfields can be grouped into three main categories based
on their goals:

Supervised learning
Unsupervised learning
Reinforcement learning

Let’s explore these types of machine learning.

#1: Supervised learning

Supervised learning takes place when an algorithm learns to correctly
predict responses to input data by using examples and their associated target
responses. The target responses can range from numerical values to string
labels such as tags. This learning process is like a teacher giving examples
to students in a class, who then memorize these examples and extract
general rules from them. Examples of these learning algorithms include the
following:

Decision trees
Random forest
KNN
Regression algorithms

Chapter 5 discusses these algorithms in greater detail.

#2: Unsupervised learning

As the name suggests, unsupervised learning takes place when an algorithm
learns from examples without any associated responses. In other words, the
algorithm is left to determine its own data patterns from the given set of
inputs. This type of algorithm usually restructures the data into classes or
even a new series of uncorrelated values, kind of like a person creating
categories of similar objects or events. Unsupervised learning is quite useful
for discovering new insights from large sets of data and generating new
inputs for supervised machine-learning algorithms. Examples of
unsupervised learning algorithms include the following:

Clustering algorithm
Markov algorithm
Neural networks

These algorithms are further discussed in Chapter 6.

#3: Reinforcement learning

Reinforcement learning is like unsupervised learning except with input that
is accompanied by positive or negative feedback mechanisms depending on
how the algorithm works. Essentially the equivalent of trial and error, this
type of learning is useful for applications that require the algorithm to make
its own decisions based on the consequences of its actions, seeking good
outcomes while trying to avoid mistakes. A perfect example of such an
application is Google’s DeepMind program that plays Atari games. The
following are typical reinforcement machine-learning algorithms:

Q-learning
SARSA

Chapter 7 focuses on these algorithms.

 Chapter 5

Supervised Machine-Learning Algorithms

As we had discussed earlier, supervised learning algorithms are those

algorithms that consist of a target and outcome variable which is to be
forecasted from a given set of predictors that forms the independent
variables. By using these sets of variables, you can produce a function that
maps the inputs to desired outputs. Such a training process continues until
the machine-learning model achieves its desired level of accuracy on the
training data.

Decision Trees, Regression, Random Forest, KNN and Logistic Regression

are some examples of supervised learning algorithms. We are going to
understand and learn more about them. This chapter delves deeper into the
mechanics behind the main types of supervised machine-learning
algorithms. Let’s jump in.
Decision trees
Decision trees are excellent, highly efficient data-analysis tools that can
help you choose one alternative from several alternative actions. They
provide a highly efficient structure that allow you to lay out options and
examine the possible outcomes of each option, providing you with a
balanced picture of all their pros and cons. Surprisingly, these algorithms
work for both categorical and continuous random variables, though they are
mostly used to solve classification problems. A decision tree splits the
entire domain set into two or more sets of similar data, sorting them by
independent variables that distinguish the sets. Let me illustrate using an
example.

Suppose you have a sample of 60 students that has three independent

variables: gender (Boy/Girl), class (IY/Y), and height (5 to 6 feet). You
know that 30 of these students play soccer in their free time, and you want
to design a model that can figure out the identity of each soccer player. To
solve this problem, a decision tree divides the sample into groups based on
the three variables, seeking to create homogeneous sets of students.

Decision trees use one of several algorithms to help them split a given set
into two or more subsets, producing increasingly homogeneous outcomes;
the exact algorithm chosen depends on the types of variables involved. The
following algorithms are among the most commonly used:

The Gini index. The Gini index works as follows: If two items with
the same value of a target variable are randomly selected, then they
must be in the same class and probability. The categorical value of
the target variable is marked by another variable as either success or
failure, which the algorithm converts into a binary split within the
 class. The Gini scores of the subclasses are then calculated as the
squares of their respective probabilities within the class and added
together to determine the Gini index of the class. The higher a
class’s Gini index, the greater its homogeneity.
The higher the value of the Gini index, the greater the homogeneity.
For the sub-nodes, we compute the Gini index that is denoted as the
square of probability for success and failure using the formula:
(p2+q2).
Compute the Gini for the split using the weighted Gini score of each
node of the split.

Chi-square. The chi-square algorithm is used to find out whether

any significant statistical differences exist between sets and their
respective subsets. The degree of statistical significance is
measured by summing the squares of standardized differences
between the observed and expected frequencies of the target
variable.
Reduction in variance. Reduction in variance is used for
continuous target variables. The algorithm selects the splits with the
lowest variances among the variables.

Advantages of decision trees

A decision tree’s output is easy to understand even if you don’t
have any statistical training.
The decision-tree algorithm is one of the fastest ways to identify the
most significant variables and the relationships between them,
making it ideal for data exploration.
Compared to other learning models, the algorithm doesn’t require a
lot of data cleaning.
 The algorithm can handle both numerical and categorical variables.

Disadvantages of decision trees

Decision trees tend to overfit data, so they require many variable
constraints that are tedious to design.
The algorithm is ill suited for continuous variables because it loses
information during the categorization of variables.
Random forest
Every data scientist uses the random-forest algorithm at some point, which
is why this algorithm is considered to be the panacea for all data-science
problems. A versatile machine-learning algorithm that combines several
techniques into one robust model, random forest can perform regression,
classification, and dimensional-reduction tasks, as well as handle essential
data-exploration tasks such as treating outliers and missing values. The
random-forest algorithm consists of multiple decision trees that classify
each object in the training set; using a process that resembles voting, the
forest selects the classification chosen by the majority of the trees. Here’s
how the algorithm works:

Every object in the training set is randomly generated and can be

replaced as necessary.
If there are M input variables, then a number m<M is specified and
held constant so that each tree can select m variables at random
from M.
Each tree finds the best split for m.
All the trees grow to the largest extent possible without pruning.
The forest then predicts the outcome by aggregating all the trees’
predictions and selecting either the majority consensus for
classification or the average for regression.

Let’s look at an illustration of the random-forest algorithm in action. Note

that the figures quoted in this book are for illustrative purposes only.

Suppose the algorithm’s training set contains socioeconomic data about

10,000 Kenyans out of the total population of 40,000,000 people. For the
sake of simplicity, let the model be built on 1 target variable, salary bands in
Kenyan shillings (Ksh.). In total, 5 different decision trees are built using 3
other variables: age, gender, and level of education. Here are the categorical
values of the target variable: Suppose we select salary as our variable for
the development of the model; then we have the following scenarios:

Salary Bands

Band 1: Below Ksh. 20,000

Band 2: Ksh. 20,000 to Ksh. 100,000

Band 3: Above Ksh. 100,000

The above salary bands are then split based on the 3 variables:

Variable: Age

Age Band 1 Band 2 Band 3

Below 18 years 90% 50% 40%
18 to 27 years 75% 43% 28%
28 to 40 years 51% 30% 16%
Above 40 years 34% 22% 10%
Variable: Gender

Gender Band 1 Band 2 Band 3

Male 71% 47% 70%
Female 83% 60% 45%

Variable: Level of education

Education Band 1 Band 2 Band 3

College certificate 90% 50% 40%
College diploma 75% 43% 28%
College degree 51% 30% 16%
Master’s and above 34% 22% 10%

Using the above aggregate data, the algorithm now builds a single mean
probabilistic model to predict the salary band of one of the training data that
can map the above data into a single probabilistic model, For simplicity, we
can use the mean probabilistic model to develop the unique framework for
modeling the above information. Let’s say a 30-year-old Kenyan male who
has a college degree, which produces the following result:

Variable Input Band 1 Band 2 Band 3

Age 28 to 40 years 51% 30% 16%
Gender Male 71% 47% 70%
Education College degree 51% 30% 16%
Advantages of random forest
The random-forest algorithm can solve both classification and
regression problems.
Random forest can handle colossal sets of data and thousands of
input variables with higher dimensionality.
The algorithm can estimate missing data with a sufficient level of
accuracy.
The algorithm can correct imbalances in the error rate between
datasets.

Disadvantages of random forest

In regression problems, random forest doesn’t predict beyond the
ranges that are given in the training data, so its results may not be as
accurate as those of other algorithms.
You have little control over what the algorithm does, so you can
only assume that it gives you the output you want.
KNN algorithm
The k-nearest neighbors (KNN) algorithm searches the entire training
dataset for the k most similar examples of a certain data instance, then
summarizes the results and uses them to predict the attributes of that
instance.

The model is also competitive learning in the sense that it internally uses
competition between the model elements—the data instances—to make a
predictive decision. The overarching objective is to measure the similarities
between data instances that provide the hidden data instance that is
necessary for making predictions.

A lazy learning process, the algorithm doesn’t build a model until it is asked
for a prediction, which keeps the data relevant to the task at hand. Let’s
look at a practical example to demonstrate how the model works.

Suppose we have 150 observations of flowers from 3 different species: Iris

setosa, I. versicolor, and I. virginica. The algorithm’s objective is to predict
the species of a flower based on 4 measurements: sepal length, sepal width,
petal length, and petal width. To identify the species of flower, the
algorithm selects the training set and computes the distance between any
two data instances, basing its prediction on the flower’s k closest
resemblances. We can assume that good classification accuracy is at least
above 90% correct.

Here’s the summary of how the classification will be done:

Generate the input data. The input data forms part of the training data. In
the flower example, we have 150 observations of the iris flowers from 3
different species that have different sepal lengths, sepal widths, petal
lengths and the petal widths.
Compute the distance between any two data instances (determine the kth
nearest neighbor) to determine the similarity levels.

Locate the k most similar among the neighbors.

Generate a response based on the data instances.

Summarize the accuracy of the predictions.

Advantages of the KNN algorithm

The KNN algorithm cuts through noise in the training set as its
competitive, instance-based model seeks only the data with the
most desired attributes.
Like random forest, the KNN algorithm can easily handle large
datasets and thousands of input variables with high levels of
dimensionality.

Disadvantages of the KNN algorithm

Since the algorithm analyzes all the training data before making a
prediction, its computational cost is high compared to other
machine-learning algorithms.
Like decision trees, the KNN algorithm often overfits data, so it too
requires many variable constraints that are tedious to design.
The algorithm is unsuitable for continuous variables.
Regression algorithms
Regression analysis investigates relationships between dependent or target
variables and independent or predictor variables. This technique is used for
forecasting, time-series modeling, and finding causal relationships between
variables such as reckless driving and number of traffic accidents. These
algorithms are designed to fit a curve or line to the data points as closely as
possible. Let’s look at an example illustrating how these algorithms are
used.

Suppose you want to estimate the sales growth in your company based on
current economic conditions. Using your company’s past and current
performance data, a regression algorithm shows you the growth trend. If
you know that your company is growing about twice as fast as the overall
economy, for instance, you can predict future sales growth.

There are various types of regression algorithms, among which the

following are the most common:

Linear regression
Logistic regression
Polynomial regression
Stepwise regression
Ridge regression

Let us now examine some of the properties of these regression algorithms.

#1: Linear regression algorithm

One of the simplest and most widely used regression algorithms, this
technique establishes a best-fit linear relationship between a dependent
variable (Y), which is usually continuous, and an independent variable (X),
which can be either continuous or discrete. When Y is plotted against X on
a Cartesian plane, the regression line is generated using the equation

Y = a + (b*X) + e

where a is the Y intercept and b is the slope of the line. The line is then used
to predict the value of the target variable based on the given predictor
variable.

#2: Logistic regression algorithm

An ideal algorithm for classification problems, logistic regression is used to
determine the probability of the outcome (success or failure) of a binary
dependent variable (e.g. 0 or 1, true or false, yes or no). In this form of
regression, the probability ranges from 0 to 1 and is calculated using one of
the equations

odds = p/(1-p)
ln(odds) = ln[p/(1-p)]
logit(p) = ln[p/(1-p)] = b0 + b1X1 + b2X2 + b3X3 + ... + bkXk

where p is the probability of a particular outcome of the dependent variable,

X is the independent variable, and b is the best-fit parameter.

#3: Polynomial Regression Algorithm

As the name suggests, a polynomial- regression algorithm generates a
polynomial when the dependent variable (Y) is plotted against an
independent variable (X), whose power is greater than 1, on a Cartesian
plane. The polynomial is a best-fit curve represented by the equation

Y = a + (bX2)

where a is the Y intercept and b is the best-fit parameter.

#4: Stepwise Regression algorithm
Stepwise regression is used when dealing with multiple independent
variables. In this regression technique, the independent variables are
automatically selected based on statistical values such as R-squared, t-
statistics, and the Akaike information criterion. The model then fits the
regression data by adding and dropping the covariates one at a time in
accordance with a specified criterion, which can be among the following
most common approaches:

Standard stepwise regression simply adds or removes predictors as

needed for each step of regression.
Forward selection begins with the most significant predictor and
adds one variable at each level in the model.
Backward elimination starts with all the predictors in the model and
removes the least significant variable at each stage.

#5: Ridge Regression

Ridge regression is used when a dataset suffers from multicollinearity, in
which independent variables are so highly correlated that their variances
may deviate from their observed values even though the model’s ordinary-
least-squares estimates are unbiased. By adding a given degree of bias to
the estimates, ridge regression reduces standard errors. If the model is a
linear regression, for example, the equation

Y = a + (bX)

has an error term (e) added to become

Y = a + (bX) + e
Factors to consider when selecting a regression- analysis
technique

Below are some useful factors that you should consider when
choosing a particular regression analysis method:
How do you want to explore your data? Data exploration is an
unavoidable component of developing a predictive model and
therefore should be one of the initial considerations in selecting the
right technique.
Which statistical metrics do you want to include in your learning
algorithm? You may be interested in several metrics such as the
statistical significance of the parameters, R-squared, adjusted R-
squared, and the error term, in which case stepwise regression may
be the best technique for your purposes.
How do you want to evaluate the predictive model? Cross-
validation is the best approach to assessing the model; a simple
mean squared deviation between the observed and the predicted
values gives you a good idea of the prediction accuracy.
Does the dataset have multiple independent variables? If these
variables can confound the model, then you should avoid the
automatic selection method so you don’t have to use all the
variables simultaneously.

Advantages of regression algorithms

Regression algorithms show significant relationships between
dependent variables and independent variables.
The algorithms show the impact of multiple independent variables
on a dependent variable.
 You can compare the effects of independent variables that are
measured differently, such as the impact of price changes and
promotional activities.

Disadvantages of regression algorithms

Regression algorithms can’t be applied to classification problems.

The algorithms have a tendency to overfit data, so you need to incorporate

many variable constraints, a tedious process.
 Chapter 6

Unsupervised Machine Learning Algorithms

In this chapter, we explore the ins and outs of unsupervised machine-

learning algorithms.
Clustering algorithm
The clustering algorithm is a simple way to classify any given dataset using
a particular number of clusters that are then divided into homogeneous
subgroups. Here’s how clustering works:

The algorithm selects the number of points for each main cluster.
For the sake of clarity, let these clusters be called centroids.
Each data point clusters with the centroid with which it shares the
most attributes.
Each centroid is then divided into more clusters based on
similarities between its members.
With each new set of clusters within a centroid, further division
occurs until the centroids no longer change, which is called
convergence.

The sum of the squares of the differences between data points within a
cluster and those in the rest of their centroid constitutes the value for that
cluster. Likewise, the total value of the centroids is the sum of the squared
values for all the clusters.

Advantages of the clustering algorithm

The clustering algorithm’s computational cost is low compared to
other machine-learning algorithms.
The algorithm is a very efficient classification system.

Disadvantages of the clustering algorithm

The clustering algorithm may not work for prediction problems.
If its centroids are improperly categorized, then the entire clustering
algorithm may not produce the desired results.
Markov algorithm
The Markov algorithm uses a predetermined set of rules to translate a string
of input data into another language, which makes it useful for finding
parameters that dictate the behavior of a dataset with a known range of
values. For instance, the algorithm can convert DNA sequences into
numerical outputs that are easier to analyze for underlying patterns.

Advantages of the Markov algorithm

The algorithm is ideal for learning problems with known inputs but
unspecified parameters.
The model can reveal hidden insights that other models can’t detect.

Disadvantage of the Markov algorithm

The model requires you to create a different set of rules for each
programing language.
Neural networks
Neural networks analyze and learn patterns through various layers of
neurons. Each layer looks for patterns in an image; if a pattern is detected,
the next layer is activated, and the process repeats itself until the network is
ready to predict what object the image portrays. If the prediction is correct,
the neurons strengthen their association between the patterns and the object.
Allow me to illustrate using an example.

Suppose you design a program that learns to recognize models of cars by

various features such as the car’s color and license plate. When you show
an image of a car to the neural network, the first layer of the system detects
the edges of the car, then the following layers each try to identify a specific
characteristic, ultimately factoring in small details such as wheel patterns or
even window patterns. The more features the neural network identifies, the
more accurate its prediction tends to be. If the neural network successfully
identifies the car’s model, it learns to associate these patterns with this
particular model so that they become easier to recognize with each
successive encounter.

Advantages of neural networks

Understanding neural networks doesn’t require much formal
statistical training.
The algorithms can implicitly detect complex nonlinear
relationships between dependent variables and independent
variables.
The algorithms can identify all the possible interactions between
predictor variables and target variables.

Disadvantages of neural networks

 Neural networks have high computational costs.

The algorithms may not correctly analyze relationships between target

variables and predictor variables.
 Chapter 7

Reinforcement Machine Learning

Algorithms

Since reinforcement learners don’t require any expert supervision, the types
of problems that they are best suited for are usually complex. Examples of
these problems include the following:

Game playing. When playing a game, the learner finds the best
move based on various factors and the number of possible states
in the game. Reinforcement learning eliminates the need for
manually specifying all the rules.
Control problems. Unlike both supervised and unsupervised
learning, reinforcement learning can develop management
policies for complex issues such as elevator scheduling.

Two approaches can be employed in reinforcement learning:

 Indirect learning. Indirect learning estimates an explicit model of
the environment and then computes its optimal policy for the
estimated model.
Direct learning. In the direct-learning approach, the optimal policy
is learned even without first learning the particular model.

Now that you know the basic strategies of reinforcement learning, let’s
explore two of the most common algorithms.
Q-learning
Often used for temporal-difference learning, Q-learning is an off-policy
algorithm that learns an action-value function to provide the expected utility
of taking a given action in a particular state. Since the algorithm can use
any such function, the policy rule must specify how the learner selects a
course of action.

Once the action-value function has been determined, the optimal policy can
be constructed using the actions that have the highest values in each of the
states. One advantage of Q-learning is that it doesn’t need a model of the
environment to compare the expected utility of all the available actions. The
algorithm also doesn’t have to adapt to problems with stochastic transitions
and rewards.
SARSA
State-action-reward-state-action (SARSA) is an algorithm that describes a
Markov decision-process policy where the main function for updating the
q-value relies on the current state of the learner, the action that the learner
chooses, the reward that the learner gets for selecting the action, and the
state that the learner will be in after taking the action. Since the SARSA
algorithm learns the safest path to the solution, it receives a higher average
reward for any trial than the Q-learning algorithm does, even though it
doesn’t choose the optimal path.
CONCLUSION
From search engines to facial-recognition systems to driverless cars,
machine learning is revolutionizing the once-dull AI field of IT. Smart
programmers stay ahead of the curve by staying well informed about the
latest technological advances, and now you have all the basic tools you
need to join the revolution and start designing machine-learning programs
yourself.
FURTHER RESOURCES
1. https://www.sas.com/en_us/insights/analytics/machine-
learning.html
2. http://www.kdnuggets.com/2016/08/10-algorithms-machine-
learning-engineers.html
3. http://cdn.intechopen.com/pdfs/10694.pdf
4. http://disp.ee.ntu.edu.tw/~pujols/Machine%20Learning%20Tutorial
.pdf
5. http://mlss08.rsise.anu.edu.au/files/smola.pdf
6. http://www.ulb.ac.be/di/map/gbonte/mod_stoch/syl.pdf
7. http://scribd-download.com/essentials-of-machine-learning-
algorithms-with-python-and-r-
codes_58a2f7506454a7a940b1e8ec_pdf.html
8. http://machinelearningmastery.com/supervised-and-unsupervised-
machine-learning-algorithms/
9. https://www.saylor.org/site/wp-content/uploads/2011/11/CS405-
6.2.1.2-WIKIPEDIA.pdf
10. https://page.mi.fu-berlin.de/rojas/neural/chapter/K5.pdf
11. http://www.cs.upc.edu/~bejar/apren/docum/trans/09-clusterej-
eng.pdf
12. http://webee.technion.ac.il/people/shimkin/LCS11/ch4_RL1.pdf
13. http://web.mst.edu/~gosavia/neural_networks_RL.pdf
 ABOUT THE AUTHOR
Matt Gates is an associate university lecturer with more than 10 years of
teaching experience on academic subjects ranging from IT management,
software development, and machine learning to data modeling. He believes
that the next phase of IT development lies in AI, Machine Learning and
Automation.

Matt hopes to use his books to share his knowledge to impact thousands of
people.

In his spare time, Matt often engage in discussions on Reddit, forums and
buy the latest gadgets on Amazon.
Matt’s Message

Thank you for reading! This book is a starter quick guide on Machine
Learning to help you familiarize with the basic understanding on what is
Machine Learning and the various type of algorithms (including their pros
and cons) that are available for your further exploration.

If you would like to read more great books like this one, why not subscribe
to our website.

https://www.auvapress.com/vip
Thanks for reading! Please add your short review on Amazon and let me know what your thoughts! –
Matt
Other Victor’s Titles You Will Find Useful

Blockchain Technology

Blockchain is a revolution that you should not ignore anymore.

Imagine you are been presented with an opportunity before the
flourishing of Internet, what would you do? Now is the time!

You will understand everything you need to know about

the mechanics of Blockchain.
You will learn how you can benefit from Blockchain
You will learn the legal implications of Blockchain
technology

Victor Finch
ISBN: 978-1-5413-6684-8
Paperback: 102 Pages
eBook, Audiobook Available
Bitcoin

Are you still wondering or clueless about what is Bitcoin? Do you

know Bitcoin is thriving robustly as a digital currency because of its
characteristics for more than 8 years.

You will understand everything (including merits & demerits) you

need to know about Bitcoin
You will learn how to use Bitcoin and read the transactions.
You will learn discover the best practices in using Bitcoin securely.

Victor Finch
ISBN: 978-1-5441-4139-8
Paperback: 98 Pages
eBook, Audiobook Available
Other Auva Press Titles You Will Find Useful

Smart Contracts

Smart Contract is about the revolutionary (Blockchain Technology)

approach with legal contracts or any legal agreements. This book
offers an unprecedented peek into what the future may be like that
could possibly change and enhance the traditional way of doing
things for the better (many benefits).

You will learn how disruptive (positive) are Smart

Contracts
You will learn about the legal perspectives of Smart
Contracts.
BONUS Highlight: More than 7 Possible Smart Contract
Use Cases in different industries.

Victor Finch
ISBN: 978-1-5446-9150-3
Paperback: 106 Pages
eBook, Audiobook Available
Python

Python is a highly sought after skillset by many corporations.

Possibilities with Python are limitless and often prefer over Java and
C++ due to three characteristics that you will discover in this book.

You will learn how to set up your first python.

You will learn how to properly do error handling and
debug to save you hours of time.
BONUSES Included (plus Hands On Challenges)

Ronald Olsen
ISBN: 978-1-5426-6789-0
Paperback: 152 Pages
eBook, Audiobook Available
 AUVA PRESS
AUVA Press commits lots of effort in the content research, planning and production of quality books.
Every book is created with you in mind and you will receive the best possible valuable information in
clarity and accomplish your goals.
If you like what you have seen and benefited from this helpful book, we would appreciate your
honest review on Amazon or on your favorite social media.
Your review is appreciated and will go a long way to motivate us in producing more quality books
for your reading pleasure and needs.
Visit Us Online
AUVA PRESS Books
https://www.auvapress.com/books
Register for Updates
https://www.auvapress.com/vip

Contact Us
AUVA Press books may be purchased in bulk for corporate, academic, gifts
or promotional use.
For information on translation, licenses, media requests, please visit our
contact page.
https://www.auvapress.com/contact
- END -