Machine Learning
"A computer program is said to learn from experience E with respect to some
class of tasks T and performance measure P, if its performance at tasks in T,
as measured by P, improves with experience E“ – T. Michell (1997)
Example: A program for soccer tactics
• Task : Win the game
• Performance : Goals
• Experience : (x) Players’ movements (y) Evaluation
Why do Automate?
A few thousand years ago:
Manual Plowing
Today:Automated Plowing
Path of Machine Evolution…
Automation Evolution
System that Do
• Replicate repetitive human
actions
System that Think
• Cognitive capabilities handle
judgment-oriented tasks
System that
Learn/Adapt
• Learn to understand context
and adapt to users and
systemsRobotic Automation
CognitiveAutomation
IntelligentAutomation
Natural
Language
Processing
Big Data
Analytics
Artificial
Intelligence
Machine
Learning
Large Scale
Processing
Adaptive
Alteration
Rule Engine
Screen
Scraping
Workflow
Unstructured
Data
Processing
(Extraction)
Knowledge
Modelling
(Ontologies)
Implementation:
• Macro-based applets
• Screen Scraping data collection
• Workflow Implementation
• Process Mapping
• Business Process Management
Implementation:
• Built-in Knowledge repository
• Learning capabilities
• Ability to work with unstructured data
• Pattern recognition
• Reading source data manuals
Implementation:
• Artificial Intelligence Systems
• Natural Language Understanding and Generation
• Self Optimizing / Self Learning
• Predictive Analytics / hypothesis generation
• Evidence based learning
Capabilities
Capabilities
Capabilities
Evolution of Machine Intelligence
• Raw computing power can automate
complex tasks!Great Algorithms
+ Fast Computers
• Automating automobiles into autonomous
automata!More Data + Real-
Time Processing
• Automating question answering and
information retrieval!Big Data + In-
Memory Clusters
• Deep Learning + Smart Algorithms =
Master Gamer
Deep Learning
• New algorithm learns handwriting of
unseen symbols from very few training
examples (unlike typical Deep Learning)
ImproveTraining
Efficiency
IBM Deep
Blue
Google Self
Driven Cars
Watson
Jeopardy
Deepmind
Atari Game
One Shot
Learning
Why Machine Learning?
Human Behavior & their Life are not logical like Code, not linear like a Formulas and not consistent like Rules, so it is hard for Machines
to understand & respond to humans, that is the challenge for todays Digital world. Unless, Machine starts to Learn this ever changing
human behavior, it can neither understand effectively nor respond intelligently & personally with its human counterpart. Machine
Learning algorithms offers a mechanism to understand this non-linear, non-consistent and intuitive behavior.
Formula
Behavior
Actual
Behavior
Machine Learning to help machine Learn
about Human World.
Types of Tasks for ML
Decide between two classes
Group data points tightly
Fit the target values
Classification
Regression
Clustering
AnomalyDetection
Find something out of place
Calls to Customer Care
Delta Change in Calls
Duration
Grouping by distance from
tower
Call drops due to technical
issues
How to build Model?
Task : Prove Hypothesis
Experience : Nature of Training Data
Goal : Minimize Loss Function
Loss Function = | Predicted Value – Actual Value |
How to evaluate Model Performance?
Cross Validation
Major Reasons:
• Less relevant Feature
• Smaller Training Data Set
• Higher Polynomials
• High/Low Learning Rate
• High/Low Regularization Value“Underfitting”
What are Key Data Learning Algorithms?
Reinforcement
Learning
Learning from Data Paradigm
• Learning by fully
labelled Data
• Used For: Prediction,
Classification (discrete
labels), Regression (real
values)
• Learning by Data
interrelationship
• Used for: Clustering,
Probability distribution
estimation, Finding association
(in features)
• Learning by Feedback
Loop
• Used for: Decision making
(robot, chess machine)
• Learning by partially
labelled and Data
interrelationships
• Used For: Prediction,
Classification (discrete
labels), Regression (real
values)
What are Key Problem Solving Algorithms?
ProblemType Paradigm
What is probable
effect of it?
How can we generalize
given model?
Is this A or B? Is
this A or B or C?
What is its decision
flow/reasoning?
Can we draw straight
rules from it?
How is it Organized?
Can combining models gives
better output?
Classification
Algorithms
How much/How
many it is?
Can we get higher
abstraction from it?
What is common in
it?
What is the similarity
in it?
Can it draw finer
feature from it?
Is it weird? What should I do
Next?
Anomaly Detection Reinforcement
Learning
What is difference between Software vs Intelligent System Engineering?
Deployment
Monitoring Support
Testing
Regression/ Integration
System Testing
NFR / Performance Testing
Implementation
Code Implementation Unit Testing
Designing
HLD - Architecture Level LLD – Class and method level
System Analysis
Requirement Gathering
Technical Specification of
Requirements
Model Deployment
Monitoring Evaluating Managing
Model Evaluation
Error Analysis Tuning Model
Model Training
Model Selection Model Training
Feature Engineering
Feature Extraction / Processing
Feature Ranking / Selection /
Reduction
Data Preparation
Data Acquisition Data Preprocessing
Software System Engineering Process Intelligent System Engineering Process
What is NEXT in ML?
What is DL?
• “Deep Learning is a set of algorithms in Machine Learning that Attempts to model high level abstractions in data by using
architecture composed of multiple non-linear transformations.”
• Deep Learning don’t need to provide explicit Feature Engineering. It learns based on algorithm’s non learner transformation
logics.
Demo – Predicting Consumer Churn
Scenario:
• Company has been managing CRM Process for a large US based
Telecom giant.
• Lately, Client has been showing concerns about Customer churn due to
various reasons.
• Company wants to help its client by developing an Intelligent System to
predict/detect customers which are likely to abandon their
subscription.
Problem
Analysis
Data
Acquisition
Feature
Engineering
Model
Training
Model
Evaluation
State
Account
Length
Area Code Phone Int'l Plan VMail Plan
Night
Charge
Intl Mins Intl Calls
Intl
Charge
CustServ
Calls
Subscribed
(Churn)
True/False
Predicted Column
Hypothesis:
• Customer Churning can be predicted by their Usage of Calls as well as Frequency of Customer Care calls.
Objective of Demo:
• To evaluate and select best performing ML Model for predicting Customer Churn. (Build Phase)
Customer Data:
Irrelevant Columns Binary Value Columns (Yes/No)
Binary Classification Reading CSV File into
Data Frame
Removing irrelevant
columns and modifying
data value
Train models with
three best with Cross
Validation Technique
Using Confusion
Matrix – Find best
most suitable Algo