Introduction to Machine Learning Anjeli Singh Computer Science and Software Engineering April 28 th...

Preview:

Citation preview

Introduction to Machine Learning

Anjeli SinghComputer Science and Software

Engineering

April 28th 2008

Overview

• What is Machine Learning• Examples of Machine Learning

• Learning Associations• Classification• Regression• Unsupervised Learning• Reinforcement Learning• Notes

What is Machine Learning?

• We store and process data• Supermarket chain – Hundred of stores – Selling thousands of good to million of customers– Record the details: date, customer ID, goods,

money– Gigabytes of data everyday– Turn this data into information to make prediction

What is Machine Learning?

• Do we know which people are likely to buy a particular product?

• Which author to suggest to people which enjoy reading?

What is Machine Learning?

Answer is : NO

What is Machine Learning?

Answer is :NOBecause if we knew, we won’t need

any data analysis.Go ahead and write code

What is Machine Learning?

• We can collect data • Try to extract answers to these similar questions• There is a process explains the data we observe• Its not completely random E.g., Customer behavior• People don’t buy random things• E.g. when they buy beer they buy chips • There are certain pattern in the data

What is Machine Learning?

• We can’t identify process completely• We can construct good approximation• Detect certain pattern and regularities• This is the niche of Machine Learning• We can use these patterns for prediction• Assuming near future won’t be much different

the past

What is Machine Learning?

• Application to large database: data mining• Its an analogy for extracting minerals from

Earth• Large volume of data is processed – To construct a simple model with valuable use– Having a high predictive accuracy– Application :• Retail, finance banks, manufacturing, medicine for

diagnosis

What is Machine Learning?

• Its not a database problem• To be intelligent, a system that is in changing

environment should have the ability to learn• System should learn and adapt• Foresee and provide solution for all possible

situations

Mathematics Problem

• 2 + 2=

Mathematics

• 2 + 2= 4

What is Machine Learning?

• How we did that ?• Can we write a program to add two

numbers ??

Faces

• Who is He ??

Faces

Denzel Washington

What is Machine Learning?

• How do we acknowledge him ?• Can you write a program for that???

What is Machine Learning?

• Pattern Recognition Problem• Analyzing sample face images of that person• Learning captures the pattern specific to the

person• Recognize by checking this person in a given

image

What is Machine Learning?

“Machine learning is programming computers to optimize a performance criterion using example data or past experience”

• Uses theory of statistics to build mathematical models

• Core task is to make inferences from a sample

What is Machine Learning?

• Role of Computer Science– Training, need efficient algorithms to solve

optimization problems– To store and process massive data– Its representation and algorithmic solutions for

inferences– Eg, the efficiency of learning and or inference

algorithm its space and time complexity may be as imp as predictive accuracy

Examples of Machine Learning

• Learning Associations• Classification• Regression• Unsupervised Learning• Reinforcement Learning• Notes

Learning Associations

• In case of retail: Basket Analysis• Finding associations between product bought

by customers• If a customer buy X, typically also buy Y• To find Potential Y customer• Target them from cross selling

Learning Associations

• Association Rule– P(Y|X) where Y is the product we condition on X

and X is the product which a customer has already purchased

– Eg. P(chips|Beer) = 0.7 Then 70 % who buy beer also chips

– Distinction Attribute: • P(Y|X, D) D set of customer attributes• Gender, age, marital status

Classification

• Credit card Example– Predict the chances of paying loan back– Customer will default /not pay the whole amount– Bank should get profit– Not inconvenience a customer over his financial

capacity

Classification

• Credit Scoring– Calculate the risk given the amount and customer

information– Customer information. Eg., Income, savings,

profession, age, history etc.– Form a rule– Fits a Model to the past data– To calculate the risk for a new application

Classification

• Classes– Low Risk Savings

– High Risk Ѳ2

• Rule (Prediction) Ѳ1 Income

If income> Ѳ1 AND savings> Ѳ2 THEN low-risk ELSE high-risk

Example of discriminant

Low Risk

High Risk

Classification

• Decision Type– 0/1 (low-risk/high-risk)– P(Y|X) where Y customer attribute and X is 0/1– P(Y=1|X=x) =0.8

Classification

• Pattern Recognition– Optical character Recognition– Recognizing character code from images– Multiple classes– Collection of strokes, has a regularity(not random

dots)– Capture in learning a program– Sequence of characters eg . T?e word– Face recognition

Classification

• Speech recognition– Input is acoustic and classes are words

• Medical Diagnosis

Classification

• Knowledge Extraction– Learning a rule from data

• Compression– Fitting a rule to the data– Outlier Detection– Finding the instances that do not obey rule are

exceptions– E.g. Fraud

Regression• If something can predict

price of a used car?– Input: Brand, year, engine

capacity– Output: Car Price

Regression

• X denote car attribute• Y be the price• Survey past transaction • Collect training data y:price

• Fitted function• Y = wx+w0 x: milage

Supervised Learning

• There is an input and an output• Learn mapping from input to output• Model defined up to a set of parameters: y = g(x|Ѳ)

g(.) is the model and Ѳ are its parameters

y is the number of regressions or a class code (0/1)

Unsupervised Learning

• Only have input data• Aim is to find regularities in the input• There is a pattern

Unsupervised Learning

• To find the regularities in the input• Structure in the input space• Density Estimation• Clustering• Image Compression

Reinforcement Learning

Notes

• Evolution defines us• We change our behavior • To cope with change • We don’t hardwire all sort of behavior• Evolution gave us mechanism to learn• We recognize, recall the strategy• Learning has limitation also

Can we grow a third arm??

Evolution in ML

• Our aim is not understand the process underlying learning in human

• To build useful systems as in domain of engineering

Science fitting models of data

• Design experiments, observe and collect data• Extract knowledge by finding out simple

models, that explains the data• Process of extracting general rules from a set

of cases is Induction • Going from particular observation to general

description Statistics: inference Learning: estimation

Relevant Resources

• Journal of Machine Learning Research

• Neural Computation• Neural Information

Processing System (NIPS)• Book: Introduction to

Machine Learning by Ethem ALPAYDIN

The MIT Press

Recommended