Module 2: Machine Learning Deep Dive

This course content is being actively
developed by Delta Analytics, a 501(c)3
Bay Area nonprofit that aims to
empower communities to leverage their
data for good.
Please reach out with any questions or
feedback to inquiry@deltanalytics.org.
Find out more about our mission here.
Delta Analytics builds technical capacity
around the world.

Module 2:
Machine learning
building blocks.

Course overview:
Now let’s turn to the data we will be using...
✓ Module 1: Introduction to Machine Learning
✓ Module 2: Machine Learning Deep Dive
❏ Module 3: Model Selection and Evaluation
❏ Module 4: Linear Regression
❏ Module 5: Decision Trees
❏ Module 6: Ensemble Algorithms
❏ Module 7: Unsupervised Learning Algorithms
❏ Module 8: Natural Language Processing Part 1
❏ Module 9: Natural Language Processing Part 2

❏ Model Development
❏ Defining the machine learning task
❏ Measuring performance of your model
❏ Supervised vs. unsupervised learning methods
❏ Model Validation
Module Checklist

Now we have our research question, we are able to
start modeling!
Research
Question
Exploratory
Analysis
Modeling
Phase
Performance
Data
Cleaning
Learning
Methodology
Task
Research question from Module 1:
● How does loan amount requested vary by town?

Now we have our research question, we are able to
start modeling!
Research
Question
Exploratory
Analysis
Modeling
Phase
Performance
Data
Cleaning
We are here!
Learning
Methodology
Task
We will discuss model
performance in the
next module
In this module we introduce the
first two steps of modeling:
● Defining the machine learning task
● Understanding how the machine
learns

Machine learning allows us to tackle
tasks that are too difficult to code all
possible approaches to on our own.
By allowing machines to learn from
experience, we avoid the need for
humans to specify all the knowledge
a computer needs.
Let’s start at the basics. Why do we
want to build a model?
Modeling
Phase
Source: Deep Learning Book - Chapter 5: Introduction to Machine Learning

Based on our
experience of the
world, we have an
understanding of
relationships between
features
Machine learning models quantify and learn the
patterns we observe in data.
Computers acquire
human intuition and
quantify it, by extracting
patterns from raw data
Human Intuition Machine Learning
Model

Task
What is the problem we want our
model to solve?
Performance
Measure
Quantitative measure we use to
evaluate the model’s performance.
Learning
Methodology
ML algorithms can be supervised or
unsupervised. This determines the
learning methodology.
All models have 3 key components: a task, a
learning methodology and a performance measure
Source: Deep Learning Book - Chapter 5: Introduction to Machine Learning
Modeling
Phase

Task
model to solve?
Defining f(x)
What function will map our x (input)
as close as possible to the true Y
(output).
Feature
engineering
& selection
What is x? How do we decide what
explanatory features to include in
our model?
Today we are looking closer at each component of the
framework:
What assumptions does our model
make about the data? Do we have to
transform the data?
Is our f(x)
correct for
this problem?

Learning
Methodology
Is our model supervised or
unsupervised; how does that affect
the learning processing?
What is our
loss function?
Every supervised model has a loss
function it wants to minimize.
Optimization
process
How does the model minimize the
loss function.
Learning methodology: how does the model learn the
function that best maps x to the true Y?
How does
our ML
model learn?
Overview of how the model
teaches itself.

Performance
Quantitative measure we use to
evaluate the model’s performance.
Measures of
performance
R2, Adjusted
R2, MSE
Feature
Performance
Statistical
significance,
p-values
Overfitting,
underfitting,
bias, variance
Ability to
generalize to
unseen data
Performance: How do we evaluate how useful the model
is, and how we can improve it?

Supervised
learning
Unsupervised
learning
Hold on! Important disclosure:
ML
Algorithm
Labelled Data
(we have Y in our
data)
No Labelled Data
(we don’t have Y in
our data)
For the next few slides, we will be introducing the
intuition behind machine learning models using
supervised learning examples. Later in this module,
we will explore how unsupervised learning is
different.

1. Task
Task
model to solve?

Research
Question
Recap: How does loan amount
requested on Kiva vary by town in
Kenya?
We have KIVA data about
loan amount requested by
borrowers all over Kenya.
We want to know how the
loan amount requested
varies by town.

Building a model involves turning your
research question into a machine learning
question.
Task
Research question
Machine learning
task
How does loan
amount requested on
Kiva vary by town?
??

Firstly, let’s establish a common vocabulary
to talk about the data.
Source: Subset of Kenyan loans dataset.
Features
Observations
Task
Location.town is an
example of a
feature. Every
column in our data
set is a feature!
Every row of our
dataset is an
observation. When we
include the
observation in our
model it is part of
our training set.

A machine learning task has explanatory
features and an outcome feature.
Outcome featureExplanatory features
Town borrower
lives in
Loan amount
requested
Task
An outcome feature is the feature we expect to change when the explanatory
features are manipulated. In this example, we expect the loan amount to change
when we change the location.

What would the outcome features and explanatory
features be in the research questions below?
Try identifying some:
● What will the price of a stock be tomorrow?
● Does this patient have malaria?
● Would this person buy a car?

Solutions:
Explanatory features Outcome feature
Price of a stock market index today Company X’s stock price tomorrow
Age, symptoms, travel history Whether or not a patient has malaria
Income, location Whether or not a person would buy a car
The outcome feature might be
regression (e.g. $12) or a classification
(e.g. Yes or No). We’ll talk about this
more later!

Let’s define our explanatory and
outcome features for this task
Problem:
I am the mayor of a 30,000
person town and need to justify
spending budget on mosquito
nets.
I want evidence on how the
number of mosquito nets
affects the number of cases of
malaria. Can you help?
Task

Let’s start by identifying the
research question!
The research question is what
we want to find out from the
data, formally stated.
1.Research
Question
2.Task
How does the number of
cases of malaria change
when the number of
mosquito nets changes?

Next let’s define our task!
Task
Define explanatory
and outcome
feature
Define f(x)
Bring it all
together
1 2 3

explanatory feature(s) outcome feature
x Number of
mosquito nets
Number of people
with malariaY
2007: 80
2008: 40
2009: 42
2010: 35
We also call our explanatory features x, and our outcome feature Y.
Looks like as mosquito nets increase, the number of malaria cases decreases.
2007: 1000
2008: 2200
2009: 6600
2010: 12600
How does the number of cases of malaria
change when the number of mosquito nets
changes?
Define explanatory
and outcome
feature

What would you conclude from looking at this
data? How many nets would you recommend?
Task
x Number of
mosquito nets
Number of people
with malariaY
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1000
2008: 2200
2009: 6600
2010: 12600
You came to a conclusion by recognizing a pattern in the data. This is similar
to how a machine learning algorithm would approach the same problem.

Machine learning allows us to learn
from historical patterns.
If Mr. Mayor had no machine learning methods to
use, he could find an answer by trying a different #
of nets year after year.
But this has an obvious human cost, and it would
be very hard to update the model to account for,
for example, new residents to his town.
Machine learning algorithms help answer
questions without this human cost - we are
learning from data, or in other words, learning
from history!
Task

“Over four years,
increasing number of
mosquito nets
decrease the number of
malaria cases.”
● Humans form rules based upon observation and pattern
recognition.
● ML model takes input x and maps it to the output Y.
An increase in x
(mosquito nets) causes a
decrease in Y (malaria
cases).
Human Intuition Machine Learning Model

explanatory feature(s) model predicted outcome
x Number of
mosquito nets
Predicted Number of
people with malariaY*f(x) + e
Our model f(x) is a function that maps
our input x to a predicted Y*.
irreducible
error
Define f(x)

Predicted Number of
people with malariaf(x) + e = Y*
The goal of f(x) is to predict a Y* as
close to the true Y as possible.
e is irreducible error. This captures error caused by factors like measurement error,
randomness in the data, and inappropriate model choice. No matter how well you
optimize your model, this will never be reduced to 0.
My job is to make
the predictions as
useful as possible!
Our function f(x) maps an input x to
a predicted Y, which we refer to as
Y*. We want to choose an f(x) that
will map x as close to the true Y as
possible.
Define f(x)

True Y
We want Y* to be close to true Y
because we want the function to
output useful predictions.
Y*=f(x)+e
In this example,
predicted Y appears
close to the true Y.
We will talk about
how to quantify this
in the next section.
Define f(x)

True Y
We want Y* to be close to true Y because
we want the function to output useful
predictions.
Y*=f(x)+e
In this example,
predicted Y appears
far from the true.
This is probably not
very useful. We will
talk about how to
quantify this in the
next section.
Define f(x)

What is f(x)? It depends on the machine
learning algorithm we choose.
Define f(x)
x f(x) Y*
explanatory
feature(s), like
number of
mosquito nets
predicted outcome,
e.g. Number of
people with malaria
Supervised learning algorithms:
- Linear regression
- Decision tree
- Random forest
- ...
Unsupervised learning
algorithms:
- K-means clustering
- Hierarchical clustering
- ...
Examples of f(x):
When you have labelled data When you don’t have labelled data

Research question
Machine learning task
x
Number of
mosquito nets
Number of people
with malariaY*
Let’s bring everything together.
f(x)
Bring it all
together
How does the number of cases of malaria
change when the number of mosquito nets
changes?

Regression
Classification
Continuous variable
Categorical variable
A classification problem is when
we are trying to predict whether
something belongs to a category,
such as “red” or “blue” or
“disease” and “no disease”.
A regression problem is when we
are trying to predict a numerical
value, such as “dollars” or
“weight”.
Source: Andrew Ng, Stanford CS229 Machine Learning Course
The task function depends upon the data type you
want to predict. Supervised learning problems fall into
two main categories: regression & classification.
Supervised
learning task
The Task

2. Learning Methodology
Learning
Methodology
ML algorithms can be supervised or
unsupervised. This determines the
learning methodology.

Learning
Methodology
Is our model supervised or
unsupervised? How does that affect
the learning processing?
What is our
loss function?
Every supervised model has a
loss function it wants to
minimize.
Optimization
process
How does the model minimize
the loss function?
Learning methodology: how does the model learn the
function that best maps x to the true Y?
How does
our ML
model learn?
Overview of how the model
teaches itself.

Recall that machine learning is a subset of data science
that allows machines to learn from raw data.
Traditional software programing involves giving machines instructions which they perform.
Machine learning involves allowing machines to learn from raw data so that the computer
program can change when exposed to new data (learning from experience).
Source: https://www.youtube.com/watch?v=IpGxLWOIZy4
Machine learning

Machine learning is a subset of
data science that allows machines
to learn from raw data.
Learning
Methodology
What do we mean when we say a machine
“learns from experience”?
How does the
model learn from
raw data?

How the algorithm learns depends upon
type of data you have.
START:
Research question
Do I have data that may
answer my question?
Gather or find more data Do you have labelled data?
Unsupervised learning
methods
Supervised learning
methods
You are here!
YN
YN
Learning
Methodology

What does labelled data mean?
Do you have labelled data?
The outcome feature (Y)
is not recorded in the
data. You do not have a
labelled Y.
you are interested in
predicting is recorded in
the data. If you have a
labelled Y, you can use
supervised learning
methods.
Learning
Methodology
Y=Number of people
with malaria
2007: 80
2008: 40
2009: 42
2010: 35
Y=Number of people
with malaria
2007:
2008:
2009:
2010:
No
Yes

Whether or not you have labelled data
determines whether it is a Supervised or
Unsupervised Learning problem
Unsupervised Learning
● For every x, there is no Y
● Goal is not to predict, but to
investigate x
Supervised Learning
● For every x, there is a Y
● Goal is to predict Y using x
Learning
Methodology
Do you have labelled data?
Yes
No
f(x) + e = Y*
Y is in your data
Y is not in your
data

Most problems you will initially
encounter are supervised
algorithms.
How do supervised algorithms learn?

Intuitive explanation for how supervised
algorithms learn:
supervised learning
Number of people
with malaria
Y
2007: 80
2008: 40
2009: 42
2010: 35
Imagine you are a teacher and you ask
your students a question.
The labels Y provide the correct answer
for the problem the students are trying to
solve. Since you know the correct answer,
you can reward good student
performance and punish poor
performance. This encourages ongoing
learning!

Extending this example, you (the
researcher) are the teacher and the Model
is the student.
supervised learning
Model
I want to get the correct
answer for predicting Y
and be the best student in
the class.
Great Mr. Model!
Once you give me
your answer I will
let you know the
correct answer.
Every time Mr. Model predicts Y*, you compare
Y* to the true Y to see how well he did.

Our Model starts trying to provide an
estimated Y* by guessing.
supervised learning
Model
Unsurprisingly, the results appear terrible, judging from the
fact that actual numbers are very different from the predicted
numbers. To quantify how bad or good results are, we use Y-Y*.
I have never seen this
problem before! I’ll
just start by randomly
guessing an answer
and see what happens.
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40

True Y
Which model is more useful at
mapping x close to true Y?
Y*=f(x)+e
True Y
Y*=f(x)+e
Learning
Methodology
What is our
loss function?
Which prediction was worse, a) or b)?
a) b)

We can immediately tell that
b is better!
Learning
Methodology
What is our
loss function?
We can see that the f(x) in b maps x to a Y* much closer to the
true Y. A loss function allows us to quantify this difference.
True Y
Y*=f(x)+e
True Y
Y*=f(x)+e
a) b)

A loss function quantifies how
unhappy you would be if you used
f(x) to predict Y* when the
correct output is y. It is what we
want to minimize.
Another way to think about it is
that a loss function quantifies
how well our f(x) fits our data.
The Task The Loss
Function
A model’s goal is to minimize the loss function.
Source: Stanford ML Lecture 1
We have already seen one simple example: Y-Y*, or the difference between the predicted Y and the
actual Y. Later, we will see more sophisticated loss functions.

supervised learning
We first decide on how to measure how unhappy we are with
these results. We call this our loss function. On the next slide,
we show a few different possible loss functions we can use to
assess Mr. Model.
Since I know the right
answer, I can compare
predicted Y* to the
true Y to help guide
Mr. Model.
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40

Regression
Classification
Continuous variable
A classification problem is when
are trying to predict whether
something belongs to a category,
such as “red” or “blue” or
“disease” and “no disease”.
A regression problem is when we
are trying to predict a numerical
value given some input, such as
“dollars” or “weight”.
Recall that there are two different types of tasks:
Supervised
learning task
The Task

Regression
Classification
Continuous variable
The choice of loss function depends upon
the type of task. We will discuss loss
functions for both types of task.
Learning
Methodology
What is our
loss function?
log loss
absolute error
(L1)
least squares error
(L2)
hinge loss
Including mean squared
error (MSE), root mean
squared error (RMSE)

Learning
Methodology
What is our
loss function?
absolute error
(L1)
least squares error
(L2)
mean squared error
(MSE)
There are a few different loss
functions we could choose from,
depending on the problem we are
trying to solve.
log loss
hinge loss
root mean squared
error (RMSE)
Regression Classification
Continuous variable Categorical variable

Regression loss functions
1. L1 norm (mean absolute error)
2. L2 norm (least squares error)
- Mean squared error

Our outcome feature is continuous:
the number of people who have
malaria.
Learning
Methodology
What is our
loss function?
Outcome feature in
data is continuous
Number of people
with malariaY
2007: 80
2008: 40
2009: 42
2010: 35
Regression
task
L1 or L2 loss
function

L1 and L2 loss are two possible options
for assessing how unhappy we are with
Mr. Model’s choice of f(x).
Learning
Methodology
What is our
loss
function?
absolute error
(L1)
least squares error
(L2)
mean squared error
Also called L1 loss, this
minimizes the sum of
absolute errors between
True Y and predicted Y*.
Also called L2 loss, this
minimizes the square of
the error between True Y
and predicted Y*.
Takes the mean of the L2
loss over all observations.
Source: L1 and L2

What is our
loss function?
How bad were Mr. Model’s initial results?
Let’s compute the L1 norm.
Learning
Methodology
How well did Mr.
Model's random
guess perform?
Source: Intro to Stat - Introduction to Linear Regression
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40
absolute error (L1)
(|1-80|+|2000-40|+|300-42|+|40-35|)
= 2,302

What is our
loss function?
How bad where Mr. Model’s initial
results? Let’s compute the L2 norm.
Learning
Methodology
How did Mr. Model’s
initial random guess
do?
Source: Intro to Stat -Introduction to Linear Regression
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40
Least squares error
(L2)
(80-1)^2+(40-2000)^2+(42-300)^2
+(35-40)^2=3,914,430

We can normalize our L2 loss by
computing mean squared error or root
mean squared error.
Learning
Methodology
What is our
loss
function?
least squares error
(L2)
mean squared error
Also called L2 loss,
and predicted Y*.
Takes the mean of the L2
loss over all observations.
Source: L1 and L2,
root mean squared
error
Takes the square root of
the mean of the L2 loss.
RMSE = sqrt(mean(S))MSE = mean(S)

What is our
loss function?
Mean squared error takes the
average L2 error per
observation.
Learning
Methodology
How did Mr. Model’s
do?
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40
Mean squared error
((80-1)^2+(40-2000)^2+(42-300)^2
+(35-40)^2)/4
= 978,607.5
MSE = mean(S)

What is our
loss function?
Root mean squared error takes
the square root of the average
L2 error per observation.
Learning
Methodology
Predicted Number of
people with malaria
Y* Y
Actual Number of
people with malaria
2007: 80
2008: 40
2009: 42
2010: 35
2007: 1
2008: 2000
2009: 300
2010: 40
Root mean squared
error
(((80-1)^2+(40-2000)^2+(42-
300)^2+(35-40)^2)/4)^(½)
= 989.25
How did the models
do?
RMSE = sqrt(mean(S))

We can compute for each loss functions how
unhappy we are with models initial random guess.
Learning
Methodology
absolute error (L1)
least squares error
(L2)
mean squared error
minimizes the sum of
absolute errors between
True Y and predicted Y*.
and predicted Y*.
Takes the average L2 loss
per observation in the
data.
Source: L1 and L2
2,302 3,914,430 978,608
root mean squared
error
Takes the square root of
the average L2 loss per
observation in the data.
989
RMSE = sqrt(mean(S))MSE = mean(S)
Don’t worry about these numbers. What’s important is you understand how we are transforming them step by step.

This job isn’t done
until I reduce
RMSE.
There are five steps to RMSE:
Y-Y* For every observation in our
dataset, measure the difference
between true Y and predicted Y.
^2 Square each Y-Y* to get the
absolute distance, so positive
values don’t cancel out negative
ones when we sum.
Sum Sum across all observations so we
get the total error.
mean Divide the sum by the number of
observations we have.
root Take the square root of the mean
calculated above.
Model
RMSE is the square root of the average L2 loss per
observation.
MSE
Learning
Methodology

Which loss function should we use?
1. L1 norm (mean absolute error)
2. L2 norm (least squares error)

Each loss function has important pros
and cons.
Learning
Methodology
What is our
loss
function?
absolute error (L1)
least squares error
(L2)
Source: L1 and L2
Robust? Stable Solution? How many
solutions?
L1 Robust Not stable Multiple possible
solutions
L2 Not very robust Stable One possible
solution
vs.
MSE and RMSE are both normalized versions of L2
error. If we decide to use L2, we will choose MSE or
RMSE.

If we decide to use least squares
error (L2), we may decide to report
RMSE OR MSE
Learning
Methodology
What is our
loss
function?
MSE RMSEvs.
The key difference between RMSE and MSE is that taking the root in
RMSE normalizes the error to the same units of measurement.
This makes the error term more interpretable.
Both MSE and RMSE amplify and severely penalize large
errors more than small ones by squaring the error.

Classification loss functions
1. Log loss
2. Hinge loss

X Y
Temperature
Let’s define a slightly different task so
we can discuss hinge and log loss.
Define explanatory
and outcome
feature
Of patient
39.5°C
37.8°C
37.2°C
37.2°C
Does the
patient have
malaria?
No
Yes
Yes
No
Task: We want to predict whether or not a patient has malaria
using their temperature.

This is a classification task, so we can
use either log loss or hinge loss.
But first, what is a classification task?
Our outcome feature is categorical:
we want to predict whether or not
someone has malaria. This is a binary
classification problem.
Learning
Methodology
What is our
loss function?

Classification tasks output the probability of
belonging to a class. Normally, based upon a
threshold of 50% we then assign the predicted
class.
Does the model
predict that the
patient has
malaria?
Yes
Yes
Yes
No
Classify based
upon probability
threshold
outcome feature predicted probability predicted outcome
Y Y*Does the
patient have
malaria?
No
Yes
Yes
No
What is the probability
that the patient has
malaria?
0.55
0.80
0.85
0.2

We can evaluate accuracy by looking just at the
predicted outcome vs. the actual outcome. Here,
accuracy is 75%!
Does the model
predict that the
patient has
malaria?
Yes
Yes
Yes
No
Classify based
upon probability
threshold
outcome feature predicted probability predicted outcome
Y Y*Does the
patient have
malaria?
No
Yes
Yes
No
What is the probability
that the patient has
malaria?
0.55
0.80
0.85
0.2
However, we are missing out on using probability, which is important
information about how certain the model is about its prediction. Let’s look at
some loss functions that utilize this metric.

For every prediction the model Makes, we can
measure the logarithmic loss. What is logarithmic
loss?
Source: https://www.r-bloggers.com/making-sense-of-logarithmic-loss/
Log loss
I’m only 60%
sure that the
answer is
Yes…
Model
Log loss
evaluates the
probability of
belonging in a
class
- The smaller the log loss, the smaller the
uncertainty, the better the model
- A perfect classifier would have log loss = 0
- Log loss heavily penalizes classifiers that are
confident about an incorrect classification
- Ways to improve log loss:
- Are there problematic errors in dataset?
- Do we want to smooth the probabilities?

For every prediction our model Makes, we can also
measure the hinge loss. What is hinge loss?Hinge loss
Hinge loss is the logical extension of the regression loss function, absolute loss.
Absolute loss: Y- Y*, where Y and Y* are integers.
Hinge loss: max(0,1-(Y*)(Y))
Where Y can equal -1 (no) or 1 (yes) for each class.
For each observation, if Y* == Y (both are 1 or both are -1), hinge loss = 0. If Y =/= Y*, hinge loss increases.
The cumulated hinge loss is therefore the upper bound of the number of mistakes made by the classifier.
Sources: https://en.wikipedia.org/wiki/Hinge_loss;
http://scikit-learn.org/stable/modules/generated/sklearn.metrics.hinge_loss.html

How do we choose a loss function for a
classification problem?
Learning
Methodology
What is our
loss function?
Log Loss Hinge Lossvs.
Leads to more exact
probabilities, but at the
cost of accuracy
Leads to better accuracy,
but at the cost of exact
probabilities

Depends on the question you want to answer!
E.g. For a problem where we are trying to assess patient health, we know
that false positives (the model predicts you do have malaria, but you actually
don’t) are safer and generally more preferable than false negatives (the
model predicts you don’t have malaria, but you actually do.)
Therefore it is probably safer to evaluate our output as a probability of
whether or not you have malaria. We will use log loss.
How do we choose a loss function for a
classification problem?
Learning
Methodology
What is our
loss function?

We provide only a broad conceptual overview of log loss
and hinge loss as we will not be using them in our coding
lab. However, we encourage you to explore them further.
More resources can be found at the end of this module.
A note on categorical loss functions ...Learning
Methodology
What is our
loss function?

Our model computed an initial guess
using RMSE.
How does he improve on his initial
guess?

What is our
loss function?
Our initial RMSE is very high. Our model
tries a different f(x) and compares
RMSE.
Learning
Methodology
If the new f(x) reduces the loss, our
model keeps changing the f(x) in
that direction. After every change,
the model measures whether the
loss has increased, decreased or
stayed the same.
Oh no! That wasn’t
very good, let’s try
something else!
1st initial
guess
2nd update 3rd update
RMSE 1,000 1,300 800
# nets 300 100 400
As the model updates the prediction at
each step, we see that themodel is
learning - it changes in response to a
higher MSE.

Learning
Methodology
How does
our ML
model learn?
The process of changing f(x) to reduce
the loss function is called learning. It is
what makes ordinary least squares (OLS)
regression a machine learning algorithm.
For every f(x) we choose there is
an associated loss.
The learning process involves
updating f(x) in order to reach
the global minimum loss.
Our
model’s
initial
random
f(x)
gives us
a high
initial
error

Our model starts with a random f(x) and
update f(x) to make our loss as small as
possible.
Learning
Methodology
Our model’s job is to change
the parameters so that
every time he changes f(x),
the loss goes down.
Our model succeeds when he reduces the
error to its minimum.
How does
our ML
model learn?

When does our model stop?Learning
Methodology
Our model’s job is to change
the parameters so that
every time he changes f(x),
the loss goes down.
Our model succeeds when he reduces the
error to its minimum.
How does
our ML
model learn?

What if we don’t have labelled
data?
Learning
Methodology
How does
our ML
model learn?
Do you have labelled
data?
is not recorded in the
data. You do not have a
labelled Y.
you are interested in
predicting is recorded in
the data. If you have a
labelled Y, you can use
supervised learning
methods.
Y=Number of
people with
malaria
2007: 80
2008: 40
2009: 42
2010: 35
Y=Number of people
with malaria
2007:
2008:
2009:
2010:
No
Yes

When does our model stop if we
don’t have labelled data?
Learning
Methodology
How does
our ML
model learn?
Our model can update the
parameters only if we have
labelled data.
What happens if we don’t actually
have Y in our data?
We turn to unsupervised learning
techniques

Unsupervised Algorithms
● For every x, there is no Y.
● We do not know the correct answers so we
cannot act as a teacher.
● Instead, we try to get an understanding of
the distribution of x to draw inference about
Y.
Source: Machine Learning Mastery - Supervised and Unsupervised Learning
Learning
Methodology
Unsupervised learning does not have labelled data. However,
it is the most promising current area of research in
machine learning. Unlocking unsupervised learning will
fundamentally change our world.

To reach true machine intelligence, ML needs to
get better at unsupervised learning.
Humans learn mostly through unsupervised
learning: we absorb vast amounts of data from
our surroundings without needing a label.
Learning
Methodology
Why is unsupervised learning important?
Unsupervised
learning is the
cake itself
Supervised
learning is the
icing on the
cake
Yan Lecun, a deep learning researcher, made the analogy that
if intelligence was a cake, unsupervised learning would be the
cake and supervised learning would be the icing on the cake.
We know how to make the icing, but we don’t know how to
make the cake. Unsupervised learning is the holy grail of
machine learning.

Learning
Methodology
Examples of unsupervised learning: This clustering
algorithm predicts a user’s friends based upon their
activity on social networks.
Source: Quora - What is unsupervised learning with an example.
We don’t have any labelled data that tells us any
node is friends with another node (i.e., x is friends
with Y).
Instead, we can use user interactions to provide
the labels. The assumption is that if you are
interacting heavily with someone they are more
likely to be your friend.

Learning
Methodology
The clustering algorithm uses email patterns to predict
the hierarchy of a business organization.
Source: Automated Social Hierarchy Detection Through Email Network Analysis - Academic Paper
This algorithm not only takes in account
the people involved in an interaction but
also the directionality of the interaction.
Based on email
interactions, who is
the boss?

In the next class, we will
introduce model
selection and evaluation.

Let’s quickly recap what
we have learnt in this
module.

Task
Performance
Measure
What is the task and the performance measure
for our malaria net example?

Task
Accurately predict the decrease in
malaria cases for a given increase in
mosquito nets
Performance
Measure
Mean squared error
for our malaria net example?

Task
Performance
Measure
for our malaria patient example?

Task
Accurately predict whether or not a
given patient is likely to have
malaria, given his or her
temperature.
Performance
Measure
Log Loss
for our malaria patient example?

You are on fire! Go straight to the
next module here.
Need to slow down and digest? Take a
minute to write us an email about
what you thought about the course. All
feedback small or large welcome!
Email: sara@deltanalytics.org

Find out more about
Delta’s machine
learning for good
mission here.

Additional resources
Rosasco, Lorenzo, et al. "Are loss functions all the same?." Neural Computation 16.5
(2004): 1063-1076. http://web.mit.edu/lrosasco/www/publications/loss.pdf
“Loss Functions for Regression and Classification”, David Rosenberg, NYU:
https://davidrosenberg.github.io/ml2015/docs/3a.loss-functions.pdf

Module 2: Machine Learning Deep Dive

More Related Content

What's hot

Similar to Module 2: Machine Learning Deep Dive

More from Sara Hooker

Recently uploaded

In this document

Module 2: Machine Learning Deep Dive