Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... ·...

55
Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides taken partly from E. ALPAYDIN

Transcript of Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... ·...

Page 1: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

Foundations of

Machine Learning

and

Data Mining Rainer Marrone, Ralf Möller

1

Today‘s slides taken partly from E. ALPAYDIN

Page 2: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

Literature

2

Page 3: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

3

Lab Class and literature

n  Thursday, 13:15 – 14:45, ES42 M2589

n  Lab Class Fr 9:45-10:30, ES42 M2589

n  First Lab Class 11.04.2011, Check StudIP for exercise sheets.

Page 4: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

4

Why “Learn” ?

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

n  There is no need to “learn” to calculate payroll n  Learning is used when:

¨  Human expertise does not exist (navigating on planet X), ¨  Humans are unable to explain their expertise (speech

recognition) ¨  Solution changes in time (routing on a computer network) ¨  Solution needs to be adapted to particular cases (user

biometrics)

Page 5: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

5

What We Talk About When We Talk About “Learning” n  Learning general models from data of particular

examples

n  Data is cheap and abundant (data warehouses, data marts); knowledge is expensive and scarce.

n  Example in retail: Customer transactions to consumer behavior:

People who bought “Da Vinci Code” also bought “The Five People You Meet in Heaven” (www.amazon.com)

n  Build a model that is a good and useful approximation to the data.

Page 6: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

6

Data Mining

Application of machine learning methods to large databases is called “Data mining”.

n  Retail: Market basket analysis, Customer relationship management (CRM)

n  Finance: Credit scoring, fraud detection n  Manufacturing: Optimization, troubleshooting n  Medicine: Medical diagnosis n  Telecommunications: Quality of service

optimization n  Bioinformatics: Motifs, alignment n  Web mining: Search engines n  ...

Page 7: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

7

What is Machine Learning?

n  Optimize a performance criterion using example data or past experience.

n  Role of Statistics: Building mathematical models, core task is inference from a sample

n  Role of Computer science: Efficient algorithms to

¨ Solve the optimization problem

¨ Representing and evaluating the model for inference

Page 8: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

8

Sample of ML Applications

n  Learning Associations n  Supervised Learning

¨  Classification

¨  Regression

n  Unsupervised Learning

n  Reinforcement Learning

Page 9: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

9

Learning Associations

n  Basket analysis: P (Y | X ) probability that somebody who buys X also buys Y where X and Y are products/services.

Example: P ( chips | beer ) = 0.7 n  If we know more about customers or make a

distinction among them: ¨  P (Y | X, D )

where D is the customer profile (age, gender, martial status, …)

¨  In case of a Web portal, items correspond to links to be shown/prepared/downloaded in advance

Page 10: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

10

Classification

n  Example: Credit scoring

n  Differentiating between low-risk and high-risk customers from their income and savings

Discriminant: IF income > θ1 AND savings > θ2 THEN low-risk ELSE high-risk

Page 11: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

11

Classification: Applications

n  Aka Pattern recognition n  Character recognition: Different handwriting styles. n  Face recognition: Pose, lighting, occlusion (glasses, beard),

make-up, hair style n  Speech recognition: Temporal dependency.

¨  Use of a dictionary for the syntax of the language. ¨  Sensor fusion: Combine multiple modalities; eg, visual (lip image)

and acoustic for speech

n  Medical diagnosis: From symptoms to illnesses n  Brainwave understanding: From signals to “states” of thought n  Reading text: n  ...

Page 12: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

12

Character Recognition

Page 13: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

13

Face Recognition

Training examples of a person

Test images

AT&T Laboratories, Cambridge UK

Page 14: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

14

Page 15: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

15

Medical diagnosis

Page 16: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

16

Page 17: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

17

Page 18: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

18

Regression

n  Example: Price of a used plane

n  x : plane attribute

y : price y = g (x | θ )

g ( ) model, θ parameters

y = wx+w0

Page 19: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

19

Supervised Learning: Uses

n  Prediction of future cases: Use the rule to predict the output for future inputs

n  Knowledge extraction: The rule is easy to understand

n  Compression: The rule is simpler than the data it explains

n  Outlier detection: Exceptions that are not covered by the rule, e.g., fraud

Page 20: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

20

Unsupervised Learning

n  Learning “what normally happens” n  No output (we do not know the right answer) n  Clustering: Grouping similar instances n  Example applications

¨  Customer segmentation in CRM n  Company may have different marketing approaches for different

groupings of customers ¨  Image compression: Color quantization

n  Instead of using 24 bits to represent 16 million colors, reduce to 6 bits and 64 colors, if the image only uses those 64 colors

¨  Bioinformatics: Learning motifs (sequences of amino acids in proteins)

¨  Document Classification in unknown Domains.

Page 21: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

21

Reinforcement Learning

n  Learning a policy: A sequence of actions/outputs n  No supervised output but delayed reward

n  Credit assignment problem n  Game playing

n  Robot in a maze

n  Multiple agents, partial observability, ...

Page 22: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

22

An Extended Example

n  “Sorting incoming Fish on a conveyor according to species using optical sensing”

Sea bass (cheap) Species

Salmon (expensive)

Page 23: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

23

n  Problem Analysis

¨  Set up a camera and take some sample images to extract features

n  Length

n  Lightness

n  Width

n  Number and shape of fins

n  Position of the mouth, etc…

n  This is the set of all suggested features to explore for use in our classifier!

Page 24: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

24

n  Preprocessing

¨  Use a segmentation operation to isolate fishes from one another and from the background

n  Information from a single fish is sent to a feature extractor whose purpose is to reduce the data by measuring certain features

n  The features are passed to a classifier

Page 25: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

25

Page 26: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

26

n  Classification ¨  Now we need (expert) information to find features that

enables us to distinguish the species.

¨  “Select the length of the fish as a possible feature for discrimination”

Page 27: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

27

Page 28: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

28

The length is a poor feature alone! à Cost of decision

Select the lightness as a possible feature.

Page 29: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

29

Page 30: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

30

n  Threshold decision boundary and cost relationship

¨  Move our decision boundary toward smaller values of lightness in order to minimize the cost (reduce the number of sea bass that are classified salmon!)

Task of decision theory

Page 31: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

31

n  Adopt the lightness and add the width of the fish

Fish xT = [x1, x2]

Lightness Width

Page 32: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

32

Page 33: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

33

n  We might add other features that are not correlated with the ones we already have. Precaution should be taken not to reduce the performance by adding such “noisy features”

n  Ideally, the best decision boundary should be the one which provides an optimal performance such as in the following figure:

Page 34: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

34

Page 35: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

35

n  However, our satisfaction is premature because the central aim of designing a classifier is to correctly classify novel input

Issue of generalization!

Page 36: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

36

Page 37: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

37

Standard data mining life cycle n  It is an iterative process with phase dependencies n  Consists of six (6) phases:

Page 38: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

38

Phases (1)

n  Business Understanding ¨  Understand project objectives and requirements

¨  Formulation of a data mining problem definition

n  Data Understanding ¨  Data collection

¨  Evaluate the quality of the data

¨  Perform exploratory data analysis

n  Data Preparation ¨  Clean, prepare, integrate, and transform the data

¨  Select appropriate attributes and variables

Page 39: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

39

Phases (2) n  Modeling

¨  Select and apply appropriate modeling techniques

¨  Calibrate/learn model parameters to optimize results

¨  If necessary, return to data preparation phase to satisfy model's data format

n  Evaluation ¨  Determine if model satisfies objectives set in phase 1

¨  Identify business issues that have not been addressed

n  Deployment ¨  Organize and present the model to the “user”

¨  Put model into practice

¨  Set up for continuous mining of the data

Page 40: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

40

Fallacies of Data Mining (1) n  Fallacy 1: There are data mining tools that

automatically find the answers to our problem ¨  Reality: There are no automatic tools that will solve your

problems “while you wait”

n  Fallacy 2: The DM process require little human intervention ¨  Reality: The DM process require human intervention in all

its phases, including updating and evaluating the model by human experts

n  Fallacy 3: Data mining have a quick ROI ¨  Reality: It depends on the startup costs, personnel costs,

data source costs, and so on

Page 41: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

41

Fallacies of Data Mining (2) n  Fallacy 4: DM tools are easy to use

¨  Reality: Analysts must be familiar with the model

n  Fallacy 5: DM will identify the causes to the business problem ¨  Reality: DM tools only identify patterns in your data,

analysts must identify the cause

n  Fallacy 6: Data mining will clean up a data repository automatically ¨  Reality: Sequence of transformation tasks must be defined

by an analysts during early DM phases

* Fallacies described by Jen Que Louie, President of Nautilus Systems, Inc.

Page 42: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

42

Remember n  Problems suitable for Data Mining:

¨  Require to discover knowledge to make right decisions

¨  Current solutions are not adequate

¨  Expected high-payoff for the right decisions

¨  Have accessible, sufficient, and relevant data

¨  Have a changing environment

n  IMPORTANT: ¨  ENSURE privacy if personal data is used!

¨  Not every data mining application is successful!

Page 43: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

Overview  Supervised  Learning

Page 44: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

44

Learning a Class from Examples

n  Class C of a “family car” ¨ Prediction: Is car x a family car?

¨ Knowledge extraction: What do people expect from a family car?

n  Output:

Positive (+) and negative (–) examples n  Input representation:

x1: price, x2 : engine power

Page 45: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

45

Training set X Das Bild kann nicht angezeigt werden. Dieser Computer verfügt möglicherweise über zu wenig Arbeitsspeicher, um das Bild zu öffnen, oder das Bild ist beschädigt. Starten Sie den Computer neu, und öffnen Sie dann erneut die Datei. Wenn weiterhin das rote x angezeigt wird, müssen Sie das Bild möglicherweise löschen und dann erneut einfügen.

Nt

tt ,r 1}{ == xX

⎩⎨⎧

=negative is if 0

positive is if 1

x

xr

⎥⎦

⎤⎢⎣

⎡=

2

1

x

xx

Page 46: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

46

Class C

( ) ( )2121 power engine AND price eepp ≤≤≤≤

Page 47: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

47

Hypothesis class H

⎩⎨⎧

=negative as classifies if 0

positive as classifies if 1)(

x

xx

h

hh

Error of h on H

E(h|X ) = (1 / N ) h xt( ) ≠ rt( )t=1

N

(a  ≠  b)  =  1  if  ≠,  0  otherwise

Page 48: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

48

S, G, and the Version Space

most specific hypothesis, S

most general hypothesis, G

h ∈ H, between S and G is consistent and make up the version space (Mitchell, 1997)

Page 49: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

49

Use the simpler one because n  Simpler to use

(lower computational

complexity)

n  Easier to train (lower

space complexity) n  Easier to explain

(more interpretable)

n  Generalizes better (lower

variance - Occam’s razor)

Noise and Model Complexity

Page 50: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

50

Multiple Classes, Ci i=1,...,K Nt

tt ,r 1}{ == xX

⎪⎩

⎪⎨⎧

≠∈

∈=

, if 0

if 1

ijr

jt

it

ti C

Cx

x

Train hypotheses hi(x), i =1,...,K:

( )⎪⎩

⎪⎨⎧

≠∈

∈=

, if 0

if 1

ijh

jt

it

ti C

Cx

xx

Page 51: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

51

Regression

( ) 01 wxwxg +=

( ) 012

2 wxwxwxg ++=( ) ( )[ ]∑

=

−=N

t

tt xgrN

gE1

21| X

( ) ( )[ ]∑=

+−=N

t

tt wxwrN

wwE1

20101

1|, X

{ }

( )tt

t

Nt

tt

xfrr

rx

=

ℜ∈

= =1,X

Partial  derivatives  of  E  w.r.t  w1  and  w0  and  se=ing  them  to  0  -­‐‑>  minimize  error

xwrw

xNx

Nxrrxw

t

tt

tt

10

221)(

−=

−=∑

Page 52: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

52

Model Selection & Generalization

n  Learning is an ill-posed problem; data is not sufficient to find a unique solution

n  The need for inductive bias, assumptions about H n  Generalization: How well a model performs on new

data

n  Overfitting: H more complex than C or f

n  Underfitting: H less complex than C or f

Page 53: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

53

Triple Trade-Off

n  There is a trade-off between three factors (Dietterich, 2003):

1.  Complexity of H, c (H),

2.  Training set size, N, 3.  Generalization error, E, on new data

¨  As N↑, E↓ ¨  As c (H)↑, first E↓ and then E↑

Page 54: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

54

Cross-Validation

n  To estimate generalization error, we need data unseen during training. We split the data as

¨ Training set (50%)

¨ Validation set (25%) ¨ Test (publication) set (25%)

n  Resampling when there is few data

Page 55: Foundations of Machine Learning and Data Miningmoeller/tuhh-lectures/ml-sose-14/01... · Foundations of Machine Learning and Data Mining Rainer Marrone, Ralf Möller 1 Today‘s slides

55

Dimensions of a Supervised Learner 1.  Model :

2.  Loss function:

3.  Optimization procedure:

( )θ|xg

( ) ( )( )∑ θ=θt

tt g,rLE || xX

( )X|min arg θ=θθ

E*