DATA MINING

LECTURE 12
Classification
Nearest Neighbor Classification
Support Vector Machines
Logistic Regression
Naïve Bayes Classifier
Supervised Learning
Illustrating Classification Task
Tid Attrib1 Attrib2 Attrib3 Class Learning
1 Yes Large 125K No
algorithm
2 No Medium 100K No

3 No Small 70K No

4 Yes Medium 120K No

Induction
5 No Large 95K Yes

6 No Medium 60K No

7 Yes Large 220K No Learn

8 No Small 85K Yes Model
9 No Medium 75K No

10 No Small 90K Yes

Model
10

Training Set
Apply
Tid Attrib1 Attrib2 Attrib3 Class Model
11 No Small 55K ?
12 Yes Medium 80K ?
13 Yes Large 110K ? Deduction
14 No Small 95K ?
15 No Large 67K ?
10

Test Set
NEAREST NEIGHBOR
CLASSIFICATION
Illustrating Classification Task
Tid Attrib1 Attrib2 Attrib3 Class Learning
1 Yes Large 125K No
algorithm
2 No Medium 100K No

3 No Small 70K No

4 Yes Medium 120K No

Induction
5 No Large 95K Yes

6 No Medium 60K No

7 Yes Large 220K No Learn

8 No Small 85K Yes Model
9 No Medium 75K No

10 No Small 90K Yes

Model
10

Training Set
Apply
Tid Attrib1 Attrib2 Attrib3 Class Model
11 No Small 55K ?
12 Yes Medium 80K ?
13 Yes Large 110K ? Deduction
14 No Small 95K ?
15 No Large 67K ?
10

Test Set
Instance-Based Classifiers
Set of Stored Cases • Store the training records
• Use training records to
Atr1 ……... AtrN Class predict the class label of
A unseen cases
B
B
Unseen Case
C
A Atr1 ……... AtrN

C
B
Instance Based Classifiers
• Examples:
• Rote-learner
• Memorizes entire training data and performs classification only
if attributes of record match one of the training examples exactly

• Nearest neighbor classifier

• Uses k “closest” points (nearest neighbors) for performing
classification
 Nearest Neighbor Classifiers
• Basic idea:
• “If it walks like a duck, quacks like a duck, then it’s
probably a duck”

Compute
Distance Test
Record

Training Choose k of the

Records “nearest” records
Nearest-Neighbor Classifiers
Unknown record  Requires three things
– The set of stored records
– Distance Metric to compute
distance between records
– The value of k, the number of
nearest neighbors to retrieve

 To classify an unknown record:

1. Compute distance to other
training records
2. Identify k nearest neighbors
3. Use class labels of nearest
neighbors to determine the
class label of unknown
record (e.g., by taking
majority vote)
Definition of Nearest Neighbor

X X X

(a) 1-nearest neighbor (b) 2-nearest neighbor (c) 3-nearest neighbor

K-nearest neighbors of a record x are data points

that have the k smallest distance to x
 1 nearest-neighbor
Voronoi Diagram defines the classification boundary

The area takes the

class of the green
point
Nearest Neighbor Classification
• Compute distance between two points:
• Euclidean distance

d ( p, q )   ( pi
i
 q)
i
2

• Determine the class from nearest neighbor list

• take the majority vote of class labels among the k-
nearest neighbors
• Weigh the vote according to distance
• weight factor, w = 1/d2
Nearest Neighbor Classification…
• Choosing the value of k:
• If k is too small, sensitive to noise points
• If k is too large, neighborhood may include points from
other classes

The value of k is the

complexity of the model
Nearest Neighbor Classification…
• Scaling issues
• Attributes may have to be scaled to prevent distance
measures from being dominated by one of the attributes
• Example:
• height of a person may vary from 1.5m to 1.8m
• weight of a person may vary from 90lb to 300lb
• income of a person may vary from $10K to $1M
Nearest Neighbor Classification…
• Problem with Euclidean measure:
• High dimensional data
• curse of dimensionality
• Can produce counter-intuitive results

111111111110 100000000000
vs
011111111111 000000000001
d = 1.4142 d = 1.4142

 Solution: Normalize the vectors to unit length

Nearest neighbor Classification…
• k-NN classifiers are lazy learners
• It does not build models explicitly
• Unlike eager learners such as decision trees
• Classifying unknown records are relatively
expensive
• Naïve algorithm: O(n)
• Need for structures to retrieve nearest neighbors fast.
• The Nearest Neighbor Search problem.
• Also, Approximate Nearest Neighbor Search
SUPPORT VECTOR
MACHINES
Support Vector Machines

• Find a linear hyperplane (decision boundary) that will separate the data
Support Vector Machines
B1

• One Possible Solution

Support Vector Machines

B2

• Another possible solution

Support Vector Machines

B2

• Other possible solutions

Support Vector Machines
B1

B2

• Which one is better? B1 or B2?

• How do you define better?
 Support Vector Machines
B1
Margin is defined
as the width
that the boundary
could be
increased by
before hitting a B2
data point
b21
b22

margin
b11

b12

• Find hyperplane maximizes the margin => B1 is better than B2

 Support Vector Machines
B1

 
w  x  b 0
 
  w  x  b 1
w  x  b  1

b11

  b12
  1 if w  x  b 1 2
f ( x )    Margin  
  1 if w  x  b  1 || w ||
Support Vector Machines
2
• We want to maximize: Margin 
|| w ||
 2
• Which is equivalent to minimizing: L( w) 
|| w ||
2
• But subjected to the following constraints:

if
if

• This is a constrained optimization problem

• Numerical approaches to solve it (e.g., quadratic programming)
Support Vector Machines
• What if the problem is not linearly separable?
Support Vector Machines
• What if the problem is not linearly separable?

⃗
𝑤 ⋅⃗
𝑥 + 𝑏=− 1+𝜉 𝑖

𝜉𝑖
‖𝑤‖
Support Vector Machines
• What if the problem is not linearly separable?
• Introduce slack variables
• Need to minimize:
 2
|| w ||  N k
L( w)   C   i 
2  i 1 
• Subject to:

if
if
Nonlinear Support Vector Machines
• What if decision boundary is not linear?
Nonlinear Support Vector Machines
• Transform data into higher dimensional space

Use the Kernel Trick

LOGISTIC REGRESSION
Classification via regression
• Instead of predicting the class of an record we
want to predict the probability of the class given
the record
• The problem of predicting continuous values is
called regression problem
• General approach: find a continuous function that
models the continuous points.
Example: Linear regression
• Given a dataset of the
form find a linear
function that given the
vector predicts the
value as
• Find a vector of weights
that minimizes the sum of
square errors

• Several techniques for

solving the problem.
Classification via regression
• Assume a linear classification boundary
𝑤 ⋅ 𝑥> 0
For the positive class the bigger
the value of , the further the point
is from the classification boundary,
the higher our certainty for the
membership to the positive class 𝑤 ⋅ 𝑥=0
• Define as an increasing
function of

For the negative class the smaller

the value of , the further the point
is from the classification boundary,
the higher our certainty for the
membership to the negative class
• Define as a decreasing function
of
𝑤 ⋅ 𝑥< 0
 Logistic Regression
The logistic function
1
𝑓 (𝑡)= −𝑡
1+ 𝑒
1
𝑃 ( 𝐶 +¿ ¿ 𝑥 ) = −𝑤 ⋅ 𝑥
1+𝑒
−𝑤 ⋅ 𝑥
𝑒
𝑃 ( 𝐶 −|𝑥 ) = − 𝑤⋅ 𝑥
1+ 𝑒
𝑃 ( 𝐶 +¿ ¿ 𝑥 )
log =𝑤 ⋅ 𝑥
𝑃 ( 𝐶 −| 𝑥 ) Logistic Regression: Find the
vectorthat maximizes the
Linear regression on the log-odds ratio
probability of the observed data
Logistic Regression in one dimension
Logistic regression in 2-d
Coefficients
Logistic Regression
• Produces a probability estimate for the class
membership which is often very useful.
• The weights can be useful for understanding the
feature importance.
• Works for relatively large datasets
• Fast to apply.
NAÏVE BAYES CLASSIFIER
Bayes Classifier
• A probabilistic framework for solving classification
problems
• A, C random variables
• Joint probability: Pr(A=a,C=c)
• Conditional probability: Pr(C=c | A=a)
• Relationship between joint and conditional
probability distributions
Pr(C , A) Pr(C | A) Pr( A) Pr( A | C ) Pr(C )

P ( A | C ) P (C )
• Bayes Theorem: P (C | A) 
P ( A)
 Bayesian Classifiers
• How to classify the new record X = (‘Yes’, ‘Single’, 80K)

Find the class with the highest

probability given the vector values.
Tid Refund M arital Taxable
Status Incom e Evade
Maximum A posteriori Probability
1 Yes Single 125K No
estimate:
2 No M arried 100K No
3 No Single 70K No
• Find the value c for class C that
4 Yes M arried 120K No maximizes P(C=c| X)
5 No Divorced 95K Yes
6 No M arried 60K No
How do we estimate P(C|X) for the
7 Yes Divorced 220K No
different values of C?
8 No Single 85K Yes
• We want to estimate P(C=Yes| X)
9 No M arried 75K No
• and P(C=No| X)
10 No Single 90K Yes
10
 Bayesian Classifiers
• In order for probabilities to be well defined:
• Consider each attribute and the class label as random variables
• Probabilities are determined from the data

Evade C
Event space: {Yes, No}
Tid Refund M arital Taxable P(C) = (0.3, 0.7)
Status Incom e Evade
Refund A1
1 Yes Single 125K No
2 No M arried 100K No
Event space: {Yes, No}
3 No Single 70K No
P(A1) = (0.3,0.7)
4 Yes M arried 120K No Martial Status A2
5 No Divorced 95K Yes Event space: {Single, Married, Divorced}
6 No M arried 60K No P(A2) = (0.4,0.4,0.2)
7 Yes Divorced 220K No
Taxable Income A3
8 No Single 85K Yes
Event space: R
9 No M arried 75K No
P(A3) ~ Normal(,2)
10 No Single 90K Yes
10

μ = 104:sample mean, 2=1874:sample var

Bayesian Classifiers
• Approach:
• compute the posterior probability P(C | A1, A2, …, An) using
the Bayes theorem
P ( A A  A | C ) P (C )
P (C | A A  A )  1 2 n

P( A A  A )
1 2 n

1 2 n

• Maximizing
P(C | A1, A2, …, An)
is equivalent to maximizing
P(A1, A2, …, An|C) P(C)
• The value is the same for all values of C.

• How to estimate P(A1, A2, …, An | C )?

Naïve Bayes Classifier
• Assume conditional independence among attributes
when class C is given:

• We can estimate from the data.

• New point is classified to class c if

is maximum over all possible values of C.

Example
• Record
X = (Refund = Yes, Status = Single, Income =80K)
• For the class C = ‘Evade’, we want to compute:
P(C = Yes|X) and P(C = No| X)
• We compute:
• P(C = Yes|X) = P(C = Yes)*P(Refund = Yes |C = Yes)
*P(Status = Single |C = Yes)
*P(Income =80K |C= Yes)
• P(C = No|X) = P(C = No)*P(Refund = Yes |C = No)
*P(Status = Single |C = No)
*P(Income =80K |C= No)
 How to Estimate Probabilities from Data?

Class Prior Probability:

Tid Refund M arital Taxable
Status Incom e Evade

1 Yes Single 125K No

2
3
No
No
M arried
Single
100K
70K
No
No
Nc: Number of records with
4 Yes M arried 120K No class c
5
6
No
No
Divorced
M arried
95K
60K
Yes
No
N = Number of records
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No M arried 75K No P(C = No) = 7/10
10
10 No Single 90K Yes
P(C = Yes) = 3/10
 How to Estimate Probabilities from Data?
Discrete attributes:

: number of instances
Tid Refund M arital
Status
Taxable
Incom e Evade having attribute and belong
1 Yes Single 125K No to class
: number of instances of
2 No M arried 100K No
3 No Single 70K No
4 Yes M arried 120K No class
5 No Divorced 95K Yes
6 No M arried 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No M arried 75K No
10 No Single 90K Yes
10
 How to Estimate Probabilities from Data?
Discrete attributes:

: number of instances
Tid Refund M arital
Status
Taxable
Incom e Evade having attribute and belong
1 Yes Single 125K No to class
: number of instances of
2 No M arried 100K No
3 No Single 70K No
4 Yes M arried 120K No class
5 No Divorced 95K Yes
6 No M arried 60K No
7 Yes Divorced 220K No
P(Refund = Yes | No) = 3/7
8 No Single 85K Yes
9 No M arried 75K No
10 No Single 90K Yes
10
 How to Estimate Probabilities from Data?
Discrete attributes:

: number of instances
Tid Refund M arital
Status
Taxable
Incom e Evade having attribute and belong
1 Yes Single 125K No to class
: number of instances of
2 No M arried 100K No
3 No Single 70K No
4 Yes M arried 120K No class
5 No Divorced 95K Yes
6 No M arried 60K No
7 Yes Divorced 220K No
P(Refund = Yes|Yes) = 0
8 No Single 85K Yes
9 No M arried 75K No
10 No Single 90K Yes
10
 How to Estimate Probabilities from Data?
Discrete attributes:

: number of instances
Tid Refund M arital
Status
Taxable
Incom e Evade having attribute and belong
1 Yes Single 125K No to class
: number of instances of
2 No M arried 100K No
3 No Single 70K No
4 Yes M arried 120K No class
5 No Divorced 95K Yes
6 No M arried 60K No
7 Yes Divorced 220K No
P(Status=Single|No) = 2/7
8 No Single 85K Yes
9 No M arried 75K No
10 No Single 90K Yes
10
 How to Estimate Probabilities from Data?
Discrete attributes:

: number of instances
Tid Refund M arital
Status
Taxable
Incom e Evade having attribute and belong
1 Yes Single 125K No to class
: number of instances of
2 No M arried 100K No
3 No Single 70K No
4 Yes M arried 120K No class
5 No Divorced 95K Yes
6 No M arried 60K No
7 Yes Divorced 220K No
P(Status=Single|Yes) = 2/3
8 No Single 85K Yes
9 No M arried 75K No
10 No Single 90K Yes
10
 How to Estimate Probabilities from Data?
• Normal distribution:
( a   ij ) 2

1 2 ij2
Tid Refund M arital Taxable P( Ai a | c j )  e
Status Incom e Evade
2 2
ij
1 Yes Single 125K No
2 No M arried 100K No • One for each pair
3 No Single 70K No
4 Yes M arried 120K No • For Class=No
5 No Divorced 95K Yes
6 No M arried 60K No
• sample mean μ = 110
7 Yes Divorced 220K No
• sample variance σ2= 2975
8 No Single 85K Yes
9 No M arried 75K No • For Income = 80
10 No Single 90K Yes
10

( 80  110 ) 2
1 
2 ( 2975)
P ( Income 80 | No)  e 0.0062
2 (54.54)
 How to Estimate Probabilities from Data?
• Normal distribution:
( a   ij ) 2

1 2 ij2
Tid Refund M arital Taxable P( Ai a | c j )  e
Status Incom e Evade
2 2
ij
1 Yes Single 125K No
2 No M arried 100K No • One for each pair
3 No Single 70K No
4 Yes M arried 120K No • For Class=Yes
5 No Divorced 95K Yes
6 No M arried 60K No
• sample mean μ = 90
7 Yes Divorced 220K No
• sample variance σ2= 25
8 No Single 85K Yes
9 No M arried 75K No • For Income = 80
10 No Single 90K Yes
10

( 80  90 ) 2
1 
2 ( 25 )
P ( Income 80 | Yes )  e 0.01
2 (5)
Example
• Record
X = (Refund = Yes, Status = Single, Income =80K)
• We compute:
• P(C = Yes|X) = P(C = Yes)*P(Refund = Yes |C = Yes)
*P(Status = Single |C = Yes)
*P(Income =80K |C= Yes)
= 3/10* 0 * 2/3 * 0.01 = 0
• P(C = No|X) = P(C = No)*P(Refund = Yes |C = No)
*P(Status = Single |C = No)
*P(Income =80K |C= No)
= 7/10 * 3/7 * 2/7 * 0.0062 = 0.0005
Example of Naïve Bayes Classifier
• Creating a Naïve Bayes Classifier, essentially
means to compute counts:
Total number of records: N = 10

Class No: Class Yes:

Number of records: 7 Number of records: 3
Attribute Refund: Attribute Refund:
Yes: 3 Yes: 0
No: 4 No: 3
Attribute Marital Status: Attribute Marital Status:
Single: 2 Single: 2
Divorced: 1 Divorced: 1
Married: 4 Married: 0
Attribute Income: Attribute Income:
mean: 110 mean: 90
variance: 2975 variance: 25
 Example of Naïve Bayes Classifier
Given a Test Record:

X = (Refund = Yes, Status = Single, Income =80K)

 P(X|Class=No) = P(Refund=Yes|Class=No)
 P(Single| Class=No)
 P(Income=80K|
Class=No)
= 3/7 * 2/7 * 0.0062 = 0.00075

 P(X|Class=Yes) = P(Refund=Yes| Class=Yes)

 P(Single| Class=Yes)
 P(Income=80K|
Class=Yes)
= 0 * 2/3 * 0.01 = 0

• P(No) = 0.7, P(Yes) = 0.3

Since P(X|No)P(No) > P(X|Yes)P(Yes)
Therefore P(No|X) > P(Yes|X)
Naïve Bayes Classifier
• If one of the conditional probabilities is zero, then
the entire expression becomes zero

• Laplace Smoothing:

• : number of attribute values for attribute

 Example of Naïve Bayes Classifier
Given a Test Record: With Laplace Smoothing

X = (Refund = Yes, Status = Single, Income =80K)

naive Bayes Classifier:
P(Refund=Yes|No) = 4/9  P(X|Class=No) = P(Refund=Yes|Class=No)
P(Refund=No|No) = 5/9  P(Single| Class=No)
P(Refund=Yes|Yes) = 1/5  P(Income=80K|
P(Refund=No|Yes) = 4/5
Class=No)
P(Marital Status=Single|No) = 3/10 = 4/9  3/10  0.0062 =
P(Marital Status=Divorced|No)=2/10 0.00082
P(Marital Status=Married|No) = 5/10
P(Marital Status=Single|Yes) = 3/6  P(X|Class=Yes) = P(Refund=Yes| Class=Yes)
P(Marital Status=Divorced|Yes)=2/6  P(Single| Class=Yes)
P(Marital Status=Married|Yes) = 1/6
 P(Income=80K|
For taxable income: Class=Yes)
If class=No: sample mean=110 = 1/5  3/6  0.01 = 0.001
sample variance=2975
If class=Yes: sample mean=90 • P(No) = 0.7, P(Yes) = 0.3
sample variance=25
• P(X|No)P(No) = 0.0005
Implementation details
• Computing the conditional probabilities involves
multiplication of many very small numbers
• Numbers get very close to zero, and there is a danger
of numeric instability
• We can deal with this by computing the logarithm
of the conditional probability
Naïve Bayes for Text Classification
• Naïve Bayes is commonly used for text classification
• For a document with k terms

Fraction of
documents in c
• = Fraction of terms from all documents in c that are.
Number of times
appears in all Laplace Smoothing
documents in c
Number of unique words
Total number of terms in all documents in c (vocabulary size)

• Easy to implement and works relatively well

• Limitation: Hard to incorporate additional features (beyond
words).
• E.g., number of adjectives used.
Multinomial document model
• Probability of document in class c:

• This formula assumes a multinomial distribution

for the document generation:
• If we have probabilities for events the probability of a
subset of these is

w
• Equivalently: There is an automaton spitting
words from the above distribution
 Example
News titles for Politics and Sports

Politics Sports
“Obama meets Merkel” “OSFP European basketball champion”
documents “Obama elected again” “Miami NBA basketball champion”
“Merkel visits Greece again” “Greece basketball coach?”
P(p) = 0.5 P(s) = 0.5
terms obama:2, meets:1, merkel:2, OSFP:1, european:1, basketball:3,
Vocabulary elected:1, again:2, visits:1, champion:2, miami:1, nba:1,
size: 14 greece:1 greece:1, coach:1
Total terms: 10 Total terms: 11

New title: X = “Obama likes basketball”

P(Politics|X) ~ P(p)*P(obama|p)*P(likes|p)*P(basketball|p)
= 0.5 * 3/(10+14) *1/(10+14) * 1/(10+14) = 0.000108

P(Sports|X) ~ P(s)*P(obama|s)*P(likes|s)*P(basketball|s)
= 0.5 * 1/(11+14) *1/(11+14) * 4/(11+14) = 0.000128
Naïve Bayes (Summary)
• Robust to isolated noise points

• Handle missing values by ignoring the instance during

probability estimate calculations

• Robust to irrelevant attributes

• Independence assumption may not hold for some

attributes
• Use other techniques such as Bayesian Belief Networks (BBN)

• Naïve Bayes can produce a probability estimate, but it is

usually a very biased one
• Logistic Regression is better for obtaining probabilities.
Generative vs Discriminative models
• Naïve Bayes is a type of a generative model
• Generative process:
• First pick the category of the record
• Then given the category, generate the attribute values from the
distribution of the category

C
• Conditional independence given C

𝐴1 𝐴2 𝐴𝑛

• We use the training data to learn the distribution

of the values in a class
Generative vs Discriminative models
• Logistic Regression and SVM are discriminative
models
• The goal is to find the boundary that discriminates
between the two classes from the training data

• In order to classify the language of a document,

you can
• Either learn the two languages and find which is more
likely to have generated the words you see
• Or learn what differentiates the two languages.
SUPERVISED LEARNING
Learning
• Supervised Learning: learn a model from the data
using labeled data.
• Classification and Regression are the prototypical
examples of supervised learning tasks. Other are
possible (e.g., ranking)
• Unsupervised Learning: learn a model – extract
structure from unlabeled data.
• Clustering and Association Rules are prototypical
examples of unsupervised learning tasks.
• Semi-supervised Learning: learn a model for the
data using both labeled and unlabeled data.
Supervised Learning Steps
• Model the problem
• What is you are trying to predict? What kind of optimization function do
you need? Do you need classes or probabilities?
• Extract Features
• How do you find the right features that help to discriminate between the
classes?
• Obtain training data
• Obtain a collection of labeled data. Make sure it is large enough,
accurate and representative. Ensure that classes are well represented.
• Decide on the technique
• What is the right technique for your problem?
• Apply in practice
• Can the model be trained for very large data? How do you test how
you do in practice? How do you improve?
Modeling the problem
• Sometimes it is not obvious. Consider the
following three problems
• Detecting if an email is spam
• Categorizing the queries in a search engine
• Ranking the results of a web search
Feature extraction
• Feature extraction, or feature engineering is the most
tedious but also the most important step
• How do you separate the players of the Greek national team
from those of the Swedish national team?

• One line of thought: throw features to the classifier and

the classifier will figure out which ones are important
• More features, means that you need more training data
• Another line of thought: Feature Selection: Select
carefully the features using various functions and
techniques
• Computationally intensive
Training data
• An overlooked problem: How do you get labeled
data for training your model?
• E.g., how do you get training data for ranking?
• Chicken and egg problem
• Usually requires a lot of manual effort and domain
expertise and carefully planned labeling
• Results are not always of high quality (lack of expertise)
• And they are not sufficient (low coverage of the space)
• Recent trends:
• Find a source that generates the labeled data for you.
• Crowd-sourcing techniques
Dealing with small amount of labeled data
• Semi-supervised learning techniques have been developed
for this purpose.

• Self-training: Train a classifier on the data, and then feed

back the high-confidence output of the classifier as input

• Co-training: train two “independent” classifiers and feed the

output of one classifier as input to the other.

• Regularization: Treat learning as an optimization problem

where you define relationships between the objects you want
to classify, and you exploit these relationships
• Example: Image restoration
Technique
• The choice of technique depends on the problem
requirements (do we need a probability
estimate?) and the problem specifics (does
independence assumption hold? do we think
classes are linearly separable?)
• For many cases finding the right technique may
be trial and error
• For many cases the exact technique does not
matter.
Big Data Trumps Better Algorithms
• If you have enough data then the algorithms
are not so important

• The web has made this

possible.
• Especially for text-related
tasks
• Search engine uses the
collective human
intelligence

Google lecture:
Theorizing from the Data
Apply-Test
• How do you scale to very large datasets?
• Distributed computing – map-reduce implementations of
machine learning algorithms (Mahaut, over Hadoop)

• How do you test something that is running

online?
• You cannot get labeled data in this case
• A/B testing

• How do you deal with changes in data?

• Active learning