Introduction to Data Science

Lecture as part of TERI-NORCE Research School //

Funded by the Norwegian Ministry of Foreign Affairs

Dr. Michel d. S. Mesquita1,2,3

LinkedIn QR code

1
NORCE Norwegian Research Centre, Norway / 2Bjerknes Centre for
Climate Research, Norway / 3M2Lab.org, Norway
Main reference //
This lecture is based on the book shown below
‘KT2018’ hereafter It provides an accessible and concise introduction to
data science.

Kelleher and Tierney (2018) Data Science. Boston: The

MIT Press Essential Knowledge Series. URL:
https://mitpress.mit.edu/books/data-science

TERI-NORCE CRS, October 2019 - New Delhi 2

Definitions //

Data science /
«… a set of principles, problem definitions,
algorithms, and processes for extracting nonobvious
and useful patterns from large data sets»

Machine learning (ML) /

«… focuses on the design and evaluation of
algorithms for extracting patterns from data»

Data mining /
«… generally deals with the analysis of structured
data and often implies and emphasis on commercial
applications»
KT2018

TERI-NORCE CRS, October 2019 - New Delhi 3

3
When to use data science? //
It has to be useful

When we deal with a large number of data examples or the

patterns are too complex for humans to discover and
extract manually

Actionable insight /

Insight = the pattern should give us relevant

information about the problem that is not obvious

Actionable = the insight we get should be something

we can use in some way

TERI-NORCE CRS, October 2019 - New Delhi 4

3
Brief history of data science //

Data science starts in the 1990s drawing knowledge from two historical
fields: data collection and data analysis

Data gathering /

Earlier marks of solstices and sunrise

Transactional data: advent of writing brought record keeping, e.g. Mesopotamia ~3200 BC

Nontransactional data: demographic data, e.g. earliest census in pharaonic Egypt ~3000
BC

Relational data model: Edgar F. Codd in 1970 publishes a paper on a model, which lays
the foundation for modern databases, such as the structure query language (SQL)

Big data: today we have volume, variety, velocity; NoSQL databases

TERI-NORCE CRS, October 2019 - New Delhi 5

Brief history of data science //
Thomas Bayes
Data science starts in the 1990s drawing knowledge from two historical (Wikipedia)
fields: data collection and data analysis

Data analysis /

Statistics: the branch of science that deals with the collection and analysis of data

Probability theory: 17th and 18th centuries, e.g. work of Blaise Pascal, Jakob
Bernoulli, Thomas Bayes, and others. Probability distributions became the new tool to
move beyond descriptive statistics

Statistical learning and modern data science: 19th century; new developments in
probability theory facilitated statistical learning; e.g. Works by Pierre Laplace, Carl
Friedrich Gauss, among others; method of least squares, which laid the foundation for
linear regression and later artificial neural network

Data visualisation and exploratory data analysis: 1780-1820 William Playfair invented
statistical graphics, such as the line chart, area chart, bar chart, pie chart; these
led the foundation for modern methods, such as the t-distributed stochastic neighbord
TERI-NORCE CRS, October 2019 - New Delhi 6
embedding (t-SNE) algorithm
Brief history of data science //
The role of women
The Countess of Lovelace
(Wikipedia)

Ada Lovelace's notes were labelled alphabetically from A

to G. In note G, she describes an algorithm for the
Analytical Engine to compute Bernoulli numbers. It is
considered to be the first published algorithm ever
specifically tailored for implementation on a computer,
and Ada Lovelace has often been cited as the first
computer programmer for this reason.
(Wikipedia)
TERI-NORCE CRS, October 2019 - New Delhi 7
Brief history of data science //
Claude Shannon
Data science starts in the 1990s drawing knowledge from two historical (Wikipedia)
fields: data collection and data analysis

Data analysis (cont.) /

20th century

Karl Pearson developed modern hypothesis testing; R. A. Fisher developed statistical

methods for multivariate analysis and maximum likelihood estimate

Alan Turing in the Second World War led the invention of the electronic computer

Warren McCulloch and Walter Pitts in 1943 proposed the first mathematical model of a
neural network

Claude Shannon in 1948 published «A Mathematical Theory of Communication» and founded

information theory

TERI-NORCE CRS, October 2019 - New Delhi 8

Brief history of data science //
Alan Turing Alan Turing
(Wikipedia)

Turing (1936, 1950)

Papers by Alan Turing on the topics of computable
numbers and artificial intelligence.

TERI-NORCE CRS, October 2019 - New Delhi 9

Brief history of data science //
Nils Nilsson
Data science starts in the 1990s drawing knowledge from two historical (Wikipedia)
fields: data collection and data analysis

Data analysis (cont.) /

Evelyn Fix and Joseph Hodges proposed a model for discriminatory analysis
(classification or pattern-recognition problem), which became the basis for modern
nearest-neighbor models

Establishment of the field of artificial intelligence at a workshop in Dartmouth

College

The term machine learning was beginning to be used to describe programs that gave a
computer the ability to learn from data

Nils Nilsson’s book Learning Machines in 1965 showed how neural networks could be used
to learn linear models for classification

TERI-NORCE CRS, October 2019 - New Delhi 10

Brief history of data science //
Ensemble fatigue
Data science starts in the 1990s drawing knowledge from two historical (Benestad et al., 2017,
fields: data collection and data analysis Nature Climate Change)

Data analysis (cont.) /

Earl Hunt, Janet Marin, and Philip Stone in 1966 developed the concept-learning system
framework, which was the progenitor of an important family of ML algorithms that
induced decision-tree models

A number of independent researchers developed and published earlier versions of the k-

means clustering algorithm

Today: some of the most important developments include ensemble models, where
predictions are made using a set of models, and deep-learning neural networks, which
have multiple layers of neurons (with applications in machine vision, natural-language
processing,…)

TERI-NORCE CRS, October 2019 - New Delhi 11

The Emergence and Evolution of Data Science //

1990s: the term «data science» came to prominence in discussions of the need for statisticians to join with computer
scientists to bring mathematical rigor to computational analysis of large data sets

1997: C.F. Jeff Wu’s public lecture «Statistics = Data Science?»

2001: William S. Cleveland published an action plan for creating a university department
in the field of data science

Also in 2001 //
Leo Breiman published the paper «Statistical Modeling: the two cultures» /
His distinction between a statistical focus on models that explain the data versus an algorithmic focus on
models that can actually predict the data highlights a core difference between statisticans and ML researchers
TERI-NORCE CRS, October 2019 - New Delhi 12
Today //

The role of a data scientist has

become very broad

It is difficult for an individual

to master all of the skill areas

KT2018
TERI-NORCE CRS, October 2019 - New Delhi 13
What are data? //

Datum /
Or piece of information is an abstraction of a
real-world entity (person, object, or event)

The terms variable, feature, and attribute are

often used interchangeably to denote an
individual abstraction

Each entity is typically described by a number of

attributes. For instance, a book might have the
following attributes: author, title, topic,
genre, publisher, price, etc.

TERI-NORCE CRS, October 2019 - New Delhi 14

What are data? What is a data set? //

Data set /
It consists of the data relating to a collection
of entitites, with each entity described in terms
of a set of attributes

In its most basic form a data set is organized in

an n*m data matrix called the analytics record
KT2018

TERI-NORCE CRS, October 2019 - New Delhi 15

Challenges //

Choice of attributes /
How do you choose the most appropriate
attributes?

Different attribute types /

Numeric, nominal and ordinal.

The data type of an attribute affects the methods we can use

to analyze and understand the data, including both the basic
statistics we can use to describe the distribution of values
that an attribute takes and the more complex algorithms we
use to ideentify the patterns of relationships between
attributes

TERI-NORCE CRS, October 2019 - New Delhi 16

Perspectives on data //

Structured data Unstructured data

Can be stored in a table, and every Each instance in the data set may
instance in the table has the same have its own internal structure, and
structure (i.e., set of attributes) this structure is not necessarily
the same in every instance

TERI-NORCE CRS, October 2019 - New Delhi 17

CRISP-DM Model //

TERI-NORCE CRS, October 2019 - New Delhi KT2018 18

Data Science ecosystem //

TERI-NORCE CRS, October 2019 - New Delhi 19

KT2018
Machine learning (ML) //

ML involves a two step process /

Identify useful patterns given some data
(modelling). Some methods include:
- Decision trees
- Regression models
- Neural networks

Once a model is created, it is used for analysis

TERI-NORCE CRS, October 2019 - New Delhi 20

Supervised versus Unsupervised Learning //

Supervised learning / Unsupervised learning /

Learn a function that maps from the There is no target attribute. The
values of the attributes to the value of algorithm has the more general task of
another (target) attribute. looking for regularities in the data.

«Supervised» because each of the - Cluster analysis

instances in the data set lists both the - Euclidean distance
input values and the output (target)
value for each instance.

Each instance in the data set must be

labeled with the value of the target
attribute.

TERI-NORCE CRS, October 2019 - New Delhi 21

Methods //

TERI-NORCE CRS, October 2019 - New Delhi 22

Source: https://tinyurl.com/yydmyo8x
Activity //
Q1 and Q2

TERI-NORCE CRS, October 2019 - New Delhi 23

Simple linear regression //

Hypothetical dataset All_India_Rainfall (mm) El_Nino_Southern_Oscillation

Ten years of accumulated All India Rainfall in mm for
the monsoon season and pre-monsoon DJF mean 1030 -1.50
ENSO values for the corresponding year
950 0.50

980 0.80
Definitions /
1099 -1.80
Let xi be the value of the 1100 -1.80
explanatory variable for the ith
datapoint, and Yi the random 1140 -1.90
variable representing the response
for the same datapoint 999 0.90

950 0.85

988 0.88

TERI-NORCE CRS, October 2019 - New Delhi 24

Simple linear regression //
Assumptions /

The mean of the response variable Yi depends on the

𝐸 𝑌# = 𝛼 + 𝛽𝑥# value of xi of the explanatory variable in a linear
fashion

𝑌# = 𝛼 + 𝛽𝑥# + 𝜖# The variation of the response variable Yi about the

mean is represented by a random variable 𝜖i

𝐸 𝜖# = 0 Note that for the mean of Yi to be as equation 1

above, 𝜖i has to be zero

The variance of 𝜖i is the same for all values of the

𝐸 𝜖# ~𝑁(0, 𝜎 0) explanatory variable (error variance). Also the
different 𝜖i (i=1,2,…,n), are independent of each
other

TERI-NORCE CRS, October 2019 - New Delhi 25

Simple linear regression //
Alternative way to write the model /

𝑎𝑋 + 𝑏~𝑁(𝑎𝜇 + 𝑏, 𝑎0𝜎 0)
Property of distributions. Also, assuming that Yi is
a fixed quantity 𝛼 + 𝛽𝑥# plus a random variable
𝑁(0, 𝜎 0 ), then:

𝑌# = 𝑁(𝛼 + 𝛽𝑥# , 𝜎 0) This is the simple linear regression model

TERI-NORCE CRS, October 2019 - New Delhi 26

Ladder of powers //

Sometimes we need to transform the data. Provided that x>1, we can use
the following transformation alternatives:

Reduce the high values in a dataset relative to low values Stretch out high values
relative to low ones

TERI-NORCE CRS, October 2019 - New Delhi 27

Artificial Neural Networks //

Deeplearning.ai
Video by Andrew Ng
URL: https://youtu.be/n1l-9lIMW7E
TERI-NORCE CRS, October 2019 - New Delhi 28
Deep Learning //

MATLAB Tech Talks

Video by Shyamal Patel
URL: https://www.youtube.com/watch?v=3cSjsTKtN9M
TERI-NORCE CRS, October 2019 - New Delhi 29
Example //
Recognising objects with Convolutional Neural Networks
Michel’s solution, based on video from Deeplearning.ai

TERI-NORCE CRS, October 2019 - New Delhi 30

Discussion based on the paper by Dhar (2013) //

TERI-NORCE CRS, October 2019 - New Delhi 31

Discussion based on the paper by Dhar (2013) //
Question 1

In which ways is data

science different from
statistics?

TERI-NORCE CRS, October 2019 - New Delhi 32

Discussion based on the paper by Dhar (2013) //
Question 2

How much do you Karl Popper

(Wikipedia)

agree/disagree with Karl

Popper’s argument for
evaluating a theory and
scientific progress?

TERI-NORCE CRS, October 2019 - New Delhi 33

Discussion based on the paper by Dhar (2013) //
Question 3

What powerful capability

is machine learning able
to give us?

TERI-NORCE CRS, October 2019 - New Delhi 34

Discussion based on the paper by Dhar (2013) //
Question 4

«…simpler models are more likely

to hold up on future observations
than more complex ones, all else
being equal» (p.68)

How much do you

agree/disagree with this
statement?

TERI-NORCE CRS, October 2019 - New Delhi 35

Discussion based on the paper by Dhar (2013) //
Question 5

Why are problem

formulation skills so
important?

TERI-NORCE CRS, October 2019 - New Delhi 36

Discussion based on the paper by Dhar (2013) //
Question 6

Where do errors come from?

TERI-NORCE CRS, October 2019 - New Delhi 37

Discussion based on the paper by Dhar (2013) //
Question 7

How does the Internet

contribute to inexpensive
large-scale randomized
experiments?

TERI-NORCE CRS, October 2019 - New Delhi 38

Discussion based on the paper by Dhar (2013) //
Question 8

Why is earth science

considered to be one of
the «data-starved areas of
inquiry»?

TERI-NORCE CRS, October 2019 - New Delhi 39

Discussion based on the paper by Dhar (2013) //
Question 9

What should universities

do to improve their
students’ data science
skills?

TERI-NORCE CRS, October 2019 - New Delhi 40

Discussion based on the paper by Dhar (2013) //
Question 10

How would you like to

integrate data science in
your own research work?

TERI-NORCE CRS, October 2019 - New Delhi 41

Summary and future work //
Thank you for your attention!
Feel free to get in touch with us //

[email protected]
In conclusion // +47 4760 2340

Data science has become a growing field, thanks to many past

researchers in the field. Climate science will benefit even more
in the years to come, as more and more researchers and students
learn how to apply different algorithms to solve climate
problems.

We hope this Climate Research School can inspire you as well! In

the next days, you will be learning several tools that can
become part of your professional toolbox!

TERI-NORCE CRS, October 2019 - New Delhi 42