LEARN HOW TO
MAKE MACHINE
LEARNING WORK
Presented by Faris Hassan
Intro
Farisology
- Ai majored
- Medical imagine researcher
- Data scientist in Fave
- Pod savings (Co-founder)
Objectives
Functioning ML
1- Learning the 4 aspects of a functioning
machine learning project.
2- Understanding the underlying process for
correct learning problem diagnoses.
Covered Today
Aspects of a functioning ML
Problem statement
Data issues
Evaluation metrics
Optimisation
final notes
Summary
Analytics
DESCRIPTIVE
What has happened?
Widely spread across
industries
PREDICTIVE
predicting the future
The main insights
generator for data-
driven organizations
PRESCRIPTIVE
Recommend steps to
achieve desired state
Widely used in leading
organizations
Machine learning domain
Prescription
Are solving the right problem?
Machine Learning is an automation tool that
is aided with intelligence based on historical
events.
CUSTOMER SERVICE
There is a drop in customer satisfaction, more customers
complaining our long time handling their enquires and
often it result in the bad experience.
ML material
IS YOUR PROBLEM DEFINTION MACHINE
LEARNING USE-CASE?
TYPES OF LEARNING
SUPERVISED
You have the targeted data
of the prediction collected
over time, and or annotated.
UNSUPERVISED
You want to explore and
discover segments and
similarities of your data.
RENFORCED
Machine is expected to
learn from the environment
and model proper
behaviour.
Safety tip
How do you know your use case?
You can do your research to find similar use-
cases and learn about possible approaches
to solve the problem.
Frame the problem to naturally fit one of the
learning types.
Data issues
How to have adequate
representation?
Data is not monolithic, static, or
homogeneous. Especially when you are
working with supervised learning problem,
two thirds of your potential issues are rotted
to your data, class representation, outliers,
and others.
Insufficient Data
Bad Quality
Under/Overfitting
Sampling Bias
Irrelevancy of
Features
MAIN DATA ISSUES
To look at early
Sufficiency of data
Its not always about the size
Certain ML use-cases are sufficient with few
hundreds of examples. Insufficient data
cause increase in variance which contributes
to performance issues like overfitting.
PREDICTING SETTELMENT
COMPUTATION
Computation methods
for the same category
of case might vary.
DEAL
The case settled to be
part of a major or minor
deal offer.
OUTSIDE
INFLUENCE
Case settled under the
influence of threat,
blackmail, and or cut
loses.
Non-representative
Inaccrurate representation
The Machine Learning model can only
perform and generalise well when it's trained
on a data that reflects the natural
representation of the cases in real
environment.
Skewness
Imbalanced data where a selection of
the data is represented.
Survey empty data
In surveys you find a question that is
ignored, this make representation
incomplete.
Fitting
Generalization is the key
POOR
QUALITY
MISSING VALUES
Unrecorded to empty
observations in the data.
ERRORS
For example presence of
invalid digits or non text in
a text data.
OUTLIERS
Outliers can be handled
either by deletion or by
winsorizing.
NOISE
Irrelevant data combined
into the dataset.
IRRELEVANT FEATURES
GENERATION
Using existing features
combined with new ones to
generated new set of
features.
EXTRACTION
Extracting features that are
new, like using computer
vision tehniques to extract
image descriptors.
ML main component is features engineering which takes the following steps:
"Coming up with features is
difficult, time-consuming,
requires expert knowledge.
‘Applied machine learning’ is
basically feature engineering"
Prof. Andrew Ng
Bias/Variance Tradeoff
METRICS
How to measure success?
WHAT METRICS TO LOOK AT?
Statistics
MODEL ACCURACY
PRECISION
SENSITIVITY
MAE
Business KPIs
ROI
CR
CTR
INTANGIBLES
OPTIMIZATION
BETTER USER EXPERIENCE
Personalised and localised content plays an important
role in giving the user a unique experience tailored to the
prefrences of the user.
IMPROVE PERFORMANCE
Dynamic pricing and product recommendations are
applications of optimisation where algorithmically
products are picked and presented to increase likelihood
of conversion.
MAJOR
CHALLENGES
Managing expectations
Unrealistic expectation is by far the worst enemy
for your project.
Project management
Managing your data science or machine learning
project is similar and different than managing
software engineering projects.
Tracking and monitoring
Tracking the performance might require more
resources and expertise.
Maturity
Driving your ML to maturity is time consuming,
requires a lot of efforts and support.
Best tips
MVP FIRST
Scope down your project into a small feature, go live
and build on from there.
A/B TESTING
There is usually more than one model that performs
well in theory, A/B testing is your only friend to know
which is better for your use-case.
LEVERAGE ON TECH
CICD, containerisation, and Kubernetes are going to be
your best hits in managing the product and delivery.
Be Inspired
IMAGINATION IS MORE
IMPORTANT THAN KNOWLEDGE.
ALBERT EINSTEIN
ADDITIONAL READING
AI MACHINE LEARNING GUIDE
An executive guide to AI
10 RED FLAGS
Save your data analytics program
NOTES FROM THE AI FRONTIER
the Value of machine/deep learning
Get in Touch
We'd love to hear your thoughts
ADDRESS
127.0.0.1
PHONE
012 566 9870
EMAIL
fareshasan.ai@gmail.com

Learn How to Make Machine Learning Work

  • 1.
    LEARN HOW TO MAKEMACHINE LEARNING WORK Presented by Faris Hassan
  • 2.
    Intro Farisology - Ai majored -Medical imagine researcher - Data scientist in Fave - Pod savings (Co-founder)
  • 3.
    Objectives Functioning ML 1- Learning the4 aspects of a functioning machine learning project. 2- Understanding the underlying process for correct learning problem diagnoses.
  • 4.
    Covered Today Aspects ofa functioning ML Problem statement Data issues Evaluation metrics Optimisation final notes Summary
  • 5.
    Analytics DESCRIPTIVE What has happened? Widelyspread across industries PREDICTIVE predicting the future The main insights generator for data- driven organizations PRESCRIPTIVE Recommend steps to achieve desired state Widely used in leading organizations Machine learning domain
  • 6.
    Prescription Are solving theright problem? Machine Learning is an automation tool that is aided with intelligence based on historical events.
  • 7.
    CUSTOMER SERVICE There isa drop in customer satisfaction, more customers complaining our long time handling their enquires and often it result in the bad experience.
  • 8.
    ML material IS YOURPROBLEM DEFINTION MACHINE LEARNING USE-CASE?
  • 9.
    TYPES OF LEARNING SUPERVISED Youhave the targeted data of the prediction collected over time, and or annotated. UNSUPERVISED You want to explore and discover segments and similarities of your data. RENFORCED Machine is expected to learn from the environment and model proper behaviour.
  • 11.
    Safety tip How doyou know your use case? You can do your research to find similar use- cases and learn about possible approaches to solve the problem. Frame the problem to naturally fit one of the learning types.
  • 12.
    Data issues How tohave adequate representation? Data is not monolithic, static, or homogeneous. Especially when you are working with supervised learning problem, two thirds of your potential issues are rotted to your data, class representation, outliers, and others.
  • 13.
    Insufficient Data Bad Quality Under/Overfitting SamplingBias Irrelevancy of Features MAIN DATA ISSUES To look at early
  • 14.
    Sufficiency of data Itsnot always about the size Certain ML use-cases are sufficient with few hundreds of examples. Insufficient data cause increase in variance which contributes to performance issues like overfitting.
  • 15.
    PREDICTING SETTELMENT COMPUTATION Computation methods forthe same category of case might vary. DEAL The case settled to be part of a major or minor deal offer. OUTSIDE INFLUENCE Case settled under the influence of threat, blackmail, and or cut loses.
  • 16.
    Non-representative Inaccrurate representation The MachineLearning model can only perform and generalise well when it's trained on a data that reflects the natural representation of the cases in real environment.
  • 17.
    Skewness Imbalanced data wherea selection of the data is represented. Survey empty data In surveys you find a question that is ignored, this make representation incomplete.
  • 18.
  • 19.
    POOR QUALITY MISSING VALUES Unrecorded toempty observations in the data. ERRORS For example presence of invalid digits or non text in a text data. OUTLIERS Outliers can be handled either by deletion or by winsorizing. NOISE Irrelevant data combined into the dataset.
  • 20.
    IRRELEVANT FEATURES GENERATION Using existingfeatures combined with new ones to generated new set of features. EXTRACTION Extracting features that are new, like using computer vision tehniques to extract image descriptors. ML main component is features engineering which takes the following steps:
  • 21.
    "Coming up withfeatures is difficult, time-consuming, requires expert knowledge. ‘Applied machine learning’ is basically feature engineering" Prof. Andrew Ng
  • 22.
  • 23.
    METRICS How to measuresuccess? WHAT METRICS TO LOOK AT?
  • 24.
  • 25.
    OPTIMIZATION BETTER USER EXPERIENCE Personalisedand localised content plays an important role in giving the user a unique experience tailored to the prefrences of the user. IMPROVE PERFORMANCE Dynamic pricing and product recommendations are applications of optimisation where algorithmically products are picked and presented to increase likelihood of conversion.
  • 26.
    MAJOR CHALLENGES Managing expectations Unrealistic expectationis by far the worst enemy for your project. Project management Managing your data science or machine learning project is similar and different than managing software engineering projects. Tracking and monitoring Tracking the performance might require more resources and expertise. Maturity Driving your ML to maturity is time consuming, requires a lot of efforts and support.
  • 27.
    Best tips MVP FIRST Scopedown your project into a small feature, go live and build on from there. A/B TESTING There is usually more than one model that performs well in theory, A/B testing is your only friend to know which is better for your use-case. LEVERAGE ON TECH CICD, containerisation, and Kubernetes are going to be your best hits in managing the product and delivery.
  • 28.
    Be Inspired IMAGINATION ISMORE IMPORTANT THAN KNOWLEDGE. ALBERT EINSTEIN
  • 29.
    ADDITIONAL READING AI MACHINELEARNING GUIDE An executive guide to AI 10 RED FLAGS Save your data analytics program NOTES FROM THE AI FRONTIER the Value of machine/deep learning
  • 30.
    Get in Touch We'dlove to hear your thoughts ADDRESS 127.0.0.1 PHONE 012 566 9870 EMAIL fareshasan.ai@gmail.com