Part 9: Text Classification;

The Naïve Bayes algorithm

Francesco Ricci

Most of these slides comes from the

course:
Information Retrieval and Web Search,
Christopher Manning and Prabhakar
Raghavan
1
Content

  Introduction to Text Classification

  Bayes rule
  Naïve Bayes text classification  
Feature independence assumption  
Multivariate and Multinomial approaches
  Smoothing (avoid overfitting)  
Feature selection   Chi square and
Mutual Information   Evaluating NB
classification.

2
Standing queries

  The path from information retrieval to

text classification:   You have an
information need, say:   "Unrest in the
Niger delta region"
  You want to rerun an appropriate query
periodically to find new news items on this
topic
  Youwill be sent new documents that are
found   I.e., it’s classification not ranking
  Such queries are called standing queries  

Long used by “information professionals”   A

modern mass instantiation is Google Alerts.

4
Google alerts

5
Spam filtering: Another text
classification task
From: "" <[email protected]>
Subject: real estate is the only way... gem oalvgkay

Anyone can buy real estate with no money down

Stop paying rent TODAY !

There is no need to spend hundreds or even thousands for similar courses

I am 22 years old and I have already purchased 6 properties using the methods
outlined in this truly INCREDIBLE ebook.

Change your life NOW !

=================================================
Click Below to order:
http://www.wholesaledaily.com/sales/nmd.htm
=================================================
6
Categorization/Classification
  Given:

  A description of an instance, x ∈ X, where X

is the instance language or instance space  
Issue: how to represent text documents – the
representation determines what information is
used for solving the classification task
  A fixed set of classes:
C = {c1, c2,…, cJ}
  Determine:

  Theclass of x: c(x)∈C, where c(x) is a

classification function whose domain is X
and whose range is C   We want to know how
to build classification functions (“classifiers”).
7
 Document Classification
“planning
Test language
Data: proof
intelligence ”

(AI ) (Programming ) (HCI )

Classes:
ML Planning Semantics Garb.Coll. Multimedia GUI

Training learning planning programming garbage ... ...

Data: intelligence temporal semantics collection algorithm
reasoninglanguage memory reinforcement plan proof...
optimization network... language... region...

(Note: in real life there is often a hierarchy, not

present in the above problem statement; and also,
you get papers on "ML approaches to Garb. Coll.")

8
More Text Classification Examples

  Many search engine functionalities use classification

  Assign labels to each document or web-page:
  Labels are most often topics such as Yahoo-categories
e.g., "finance," "sports," "news>world>asia>business"
  Labels may be genres
e.g., "editorials" "movie-reviews" "news“
  Labels may be opinion on a person/product

e.g., “like”, “hate”, “neutral”

  Labels may be domain-specific

e.g., "interesting-to-me" : "not-interesting-to-me”

e.g., “contains adult language” : “doesn’t”
e.g., language identification: English, French, Chinese, …
e.g., “link spam” : “not link spam”
e.g., "key-phrase" : "not key-phrase"

9
Classification Methods (1)

  Manual classification   Used by Yahoo!

(originally; now present but
downplayed), Looksmart, about.com, ODP,
PubMed   Very accurate when job is done
by experts   Consistent when the problem
size and team
is small
  Difficult
and expensive to scale  
Means we need automatic classification
methods for big problems.

10
Classification Methods (2)   Hand-coded
rule-based systems   One technique used by CS
dept’s spam filter, Reuters, CIA, etc.   Companies
(Verity) provide “IDE” for writing such rules
  Example:assign category if document contains a
given Boolean combination of words
  Standing queries: Commercial systems have
complex query languages (everything in IR query
languages + accumulators)
  Accuracyis often very high if a rule has been
carefully refined over time by a subject expert  
Building and maintaining these rules is expensive!

11
Verity topic (a classification rule)
  Note:   maintenance
issues
(author, etc.)   Hand-

weighting of
terms
  Butit is easy to explain
the results.
Classification
Methods (3)
  Supervised learning
of a document-label
assignment function
12
  Manysystems partly rely on machine learning
(Autonomy, MSN, Verity, Enkata, Yahoo!, …)   k-
Nearest Neighbors (simple, powerful)   Naive

Bayes (simple, common method)   Support-

vector machines (new, more powerful)   … plus

many other methods
  No
free lunch: requires hand-classified training
data
  Notethat many commercial systems use a mixture
of methods.
Recall a few probability basics

  For events a and b:

13
  Bayes’ Rule

p(a,b) =p(a∩b) =p(a | b)p(b) =p(b | a)p(a)

Prio
p(b | a)p(a) p(b | a)p(a)
r
p(a | b) = =
p(b) ∑x=a,a p(b | x)p(x)
Posterior

  Odds:

p(a) p(a)
14
 O(a)= =
p(a) 1−p(a)
Bayes’ Rule Example

P(C,E) =P(C | E)P(E) =P(E |C)P(C)

P(E |C)P(C)
P(C | E) =
P(E)
P(pass exam | attend classes) = ?

= P(pass exam) * P(attend classes | pass exam)/P(attend classes)

15
= 0.7 *
Initial estimation

Correction based on a ratio

0.9/0.78
= 0.7 * 1.15 = 0.81

16
 Example explained
Pass 70% Not pass 30%

Attend 50%

Attend 90% = P(attend|pass)

Not attend 50%

Not attend 10%

P(pass) = 0.7
P(attend) = P(attend| pass)P(pass) + P(attend | not pass)P(not pass)
= 0.9*0.7 + 0.5*0.3 = 0.63 + 0.15 = 0.78 p(pass
| attend) = p(pass)*p(attend | pass)/p(attend)
17
 = 0.7 * 0.9/0.78 = 0.81

Bayesian Methods

  Our focus this lecture

  Learning and classification methods based on
probability theory
  Bayes theorem plays a critical role in
probabilistic learning and classification
  Uses prior probability of each category given
no information about an item
  Obtainsa posterior probability distribution over
the possible categories given a description of an
item.

18
 Naive Bayes Classifiers

  Task: Classify a new instance D based on a tuple

of attribute values into one of
D=x1,x2,,xn
the classes cj ∈ C

cMAP = argmaxP(c j | x1,x2,,xn )

c ∈C
j

P(x1, x2,, xn | cj )P(cj )

Maximum A
Posteriori = argmax c j∈ C
class P(x1, x2,, xn )
19
= argmaxP(x1, x2,, xn | cj )P(cj )
cj∈C

20
 Naïve Bayes Assumption
  P(cj)

  Can be estimated from the frequency of classes in

the training examples
  P(x1,x2,…,xn|cj)   O(|X|n|C|) parameters (assuming X
finite)   Could only be estimated if a very, very large
number of training examples was available – or?
  Naïve Bayes Conditional Independence Assumption:
  Assume that the probability of observing the
conjunction of attributes is equal to the product of
the individual probabilities P(xi|cj).
n

cMAP = argmaxP(cj )∏P(xi | cj ) 19 cj∈C i=1

 The Naïve Bayes Classifier
Flu

X1 X2 X3 X4 X5
runnynose sinus cough fever muscle-ache  

Conditional Independence Assumption:

features detect term presence and are
independent of each other given the class:

P(x1,, x5 |C) =P(x1 |C)•P(x2 |C)••P(x5 |C)   This

model is appropriate for binary
variables
= many
  Multivariate Bernoulli model variables

22
 =only 2 values – T or F
Learning the Model
C

X1 X2 X3 X4 X5 X6

  First attempt: maximum likelihood estimates

  simply use the frequencies in the data

Pˆ(xi | c j ) =
Pˆ(cj )= N(C=cj )
N
N(
23
 Estimated conditional
Xi =xi ,C=c j ) probability that the
attribute Xi (e.g. Fever)
N(C=c j ) has the value xi (True or
False) – we will also write
P(Fever | cj) instead of
P(Fever = T | cj)
Problem with Max Likelihood
Flu

1 2 3 4 5
X X X X X
runnynose sinus cough fever muscle-ache

P(x1,, x5 |C) =P(x1 |C)•P(x2 |C)••P(x5 |C)

  What
if we have seen no training cases where patient had flu and
muscle aches?
24
 Pˆ(X5 =T |C = flu) = N(X5 =T,C = flu) = 0
N(C = flu)
  Zeroprobabilities cannot be conditioned away, no matter the
other evidence!

ℓ= argmaxc Pˆ(c)∏i Pˆ(xi | c)

25
Smoothing to Avoid Overfitting

Pˆ(xi | c j ) = N(Xi
=xi ,C=c j ) +1 N(C=c j )
+k
# of values of X i

  Somewhat more subtle version

overall fraction in
data where Xi=x i

26
P ˆ (xi |c j ) =

N(Xi = xi,C = c j )+ mP(Xi

= xi) N(C = c j )+ m
extent of
“smoothing”
Example
docID words in document in c = China? training set 1
Chinese Beijing Chinese yes 2 Chinese Chinese Shanghai yes
3 Chinese Macao yes
4 Tokyo Japan Chinese no test set 5 Chinese
Chinese Chinese Tokyo Japan ?

Pˆ(Chinese|c) = (3+1)/(3+2) = 4/5

27
 Pˆ(Japan|c) = Pˆ(Tokyo|c) = (0+1)/(3+2) = 1/5

Pˆ(Beijing|c) = Pˆ(Macao|c) = Pˆ(Shanghai|c) = (1+1)/(3+2) = 2/5

Pˆ(Chinese|c) = (1+1)/(1+2) = 2/3

Pˆ(Japan|c) = Pˆ(Tokyo|c) = (1+1)/(1+2) = 2/3

Pˆ(Beijing|c) = Pˆ(Macao|c) = Pˆ(Shanghai|c) = (0+1)/(1+2) = 1/3

) ) )
Pˆ(c|d5) ∝ Pˆ(c · Pˆ(Chinese|c · Pˆ(Japan|c · Pˆ(Tokyo|c)

·(1 −Pˆ(Beijing|c)) · (1 −Pˆ(Shanghai|c)) · (1 −Pˆ(Macao|c))

Exercise
docID words in document in c = China? training set 1
Chinese Beijing Chinese yes 2 Chinese Chinese Shanghai yes
3 Chinese Macao yes
28
 4 Tokyo Japan Chinese no test set 5 Chinese
Chinese Chinese Tokyo Japan ?

Pˆ(Chinese|c) = (3+1)/(3+2) = 4/5

Pˆ(Japan|c) = Pˆ(Tokyo|c) = (0+1)/(3+2) = 1/5

Pˆ(Beijing|c) = Pˆ(Macao|c) = Pˆ(Shanghai|c) = (1+1)/(3+2) = 2/5

Pˆ(Chinese|c) = (1+1)/(1+2) = 2/3

Pˆ(Japan|c) = Pˆ(Tokyo|c) = (1+1)/(1+2) = 2/3

Pˆ(Beijing|c) = Pˆ(Macao|c) = Pˆ(Shanghai|c) = (0+1)/(1+2) = 1/3

Estimate the probability that the test document does not

belong to class c

29
Exercise
docID words in document in c = China? training set 1
Chinese Beijing Chinese yes 2 Chinese Chinese Shanghai yes
3 Chinese Macao yes
4 Tokyo Japan Chinese no test set 5 Chinese
Chinese Chinese Tokyo Japan ?

Pˆ(Chinese|c) = (3+1)/(3+2) = 4/5

Pˆ(Japan|c) = Pˆ(Tokyo|c) = (0+1)/(3+2) = 1/5

Pˆ(Beijing|c) = Pˆ(Macao|c) =
Pˆ(Shanghai

ˆ
P (Chinese|

Pˆ(Japan|c) = Pˆ(Tokyo|c) = (1+1)/(1+2) = 2/3

Pˆ(Beijing|c) = Pˆ(Macao|c) = Pˆ(Shanghai|c) = (0+1)/(1+2) = 1/3

30
 Pˆ
Multinomial Naive Bayes Classifiers: Basic
method

  Attributes (Xi) are text positions, values (xi) are

words:

cNB = argmaxP (xi |c j )

cj ∈C

= argmaxP(c j )P(X1 ="Our"|c j )P(Xn ="text."|c j )

 cj ∈C

  Still too many possibilities

31
  Assume that classification is independent of the
positions of the words

P(Xi =w | c) =P(X j =w | c)
Multinomial Naïve Bayes: Learning

  From training corpus, extract Vocabulary   Calculate

required P(cj) and P(xk | cj) terms   For each class cj in
C do   docsj ← subset of documents for which the target
class is cj
|docsj |
P(cj )←
total # documents

32
   Textj ← single document containing all
docsj   for each word xk in Vocabulary  

njk ← number of occurrences of xk in Textj

  nj ← number of words in Textj
Assume is = 1;
n jk +α
P(xk | cj ) ← nj +α|Vocabulary | this is for
smoothing
Multinomial Naïve Bayes:
Classifying

  positions ← all word positions in current

document which contain tokens found
in Vocabulary

  Return cNB such that:

33
 cNB = argmax P(cj ) ∏P(xi | cj ) cj∈C
i∈positions

Example
docID words in document in c = China? training set 1
Chinese Beijing Chinese yes 2 Chinese Chinese Shanghai yes
3 Chinese Macao yes
4 Tokyo Japan Chinese no test set 5 Chinese
Chinese Chinese Tokyo Japan ?

Pˆ(Chinese|c) = (5+1)/(8+6) = 6/14 = 3/7

Pˆ(Tokyo|c) = Pˆ(Japan|c) = (0+1)/(8+6) = 1/14

Pˆ(Chinese|c) = (1+1)/(3+6) = 2/9

34
 Pˆ(Tokyo|c) = Pˆ(Japan|c) = (1+1)/(3+6) = 2/9

Pˆ(c|d5) ∝ 3/4 · (3/7)3 · 1/14 · 1/14 ≈ 0.0003.

)
Pˆ(c|d5) ∝ 1/4 · (2/9 3 · 2/9 · 2/9 ≈ 0.0001.

Naive Bayes: Time Complexity

Number of
  Training
Time: if Ld is the average length of a conditional
document in D probabilities to
estimate
  O(|D|Ld + |C||V|))

Scan the documents to compute the

vocabulary and the frequencies of words

  Assumes that V and all docsj , nj, and njk are
35
 computed in O(|D|Ld) time during one pass through
all of the data
  Generally just O(|D|Ld) since usually |C||V| < |D|Ld
  TestTime: O(|C| Lt) - where Lt is the average length of
a test document
  Very efficient overall, linearly proportional to the time
needed to just read in all the data.
Underflow Prevention: log space

  Multiplying
lots of probabilities, which are between 0
and 1 by definition, can result in floating-point
underflow
  Since
log(xy) = log(x) + log(y), it is better to
perform all computations by summing logs of
probabilities rather than multiplying probabilities

36
  Classwith highest final un-normalized log probability
score is still the most probable

cNB = argmax( logP(cj) + ∑logP(x | c ) )

i j

cj∈C
i∈positions

  Note that model is now just max of sum of weights…

Sounds
familiar?
Summary - Two Models: Multivariate Bernoulli

  One feature Xw for each word in

dictionary   Xw = true in document d if w
37
appears in d   Naive Bayes
assumption:   Given the document’s
topic (class),
appearance of one word in the document tells
us nothing about chances that another word
appears (independence)
Summary - Two Models: Multinomial

  One feature Xi for each word positions in document

  feature’s values are all words in dictionary
  Value of Xi is the word in position i   Naïve
Bayes assumption:   Given the document’s
topic (class), word in one
position in the document tells us nothing about
words in other positions
38
   Second assumption:   Word appearance does
not depend on position - for
all positions i,j, word w, and class c

P(Xi =w | c) =P(X j =w | c)
  Just
have one multinomial feature predicting all
words.
Parameter estimation

  Multivariate Bernoulli model:

39
 fraction of documents of
ˆ
P (Xw = true |c j ) topic cj in which word w
appears
=
  Multinomial model:

fraction of times in which

Pˆ(Xi =w| cj ) = word w appears
across all documents of topic cj

  Can
create a mega-
document for topic j by concatenating all
documents in this topic   Use frequency of w

40
Classification

  Multinomial vs Multivariate Bernoulli?

  Multinomial model is almost always more

effective in text applications!
  See results figures later

  See
IIR sections 13.2 and 13.3 for worked
examples with each model
Feature Selection: Why?

41
  Text collections have a large number of features
  10,000 – 1,000,000 unique words … and
more
  May make using a particular classifier unfeasible
  Some classifiers can’t deal with 100,000 of
features
  Reduces training time   Training time for
some methods is quadratic or
worse in the number of features
  Can improve generalization
(performance)   Eliminates noise
features   Avoids overfitting.

42
Feature selection: how?

  Twoideas:
  Hypothesis testing statistics:   Are
we confident that the value of one
categorical variable is associated with the
value of another
  Chi-square test (
http://faculty.vassar.edu/lowry/
webtext.html chapter 8)
  Information theory:   How much
information does the value of one categorical
variable give you about the value of another
  Mutual information

43
 χ2 statistic (CHI) – testing
independence of class and term
Term = jaguar Term ≠ jaguar

Class = auto 2 500

Class ≠ auto 3 9500

  Ifthe term "jaguar" is independent from the class "auto"

we should have:   P(C(d)=auto, d contains jaguar) =
P(C(d)= auto) * P(d contains jaguar)
  2/10005=? 502/10005 * 5/10005   0.00019 =? 0.0501 *

0.00049 = 0.000025 NOT REALLY!   To be independent

we should have more documents that contains jaguar but are
not in class auto (38).
44
χ2 statistic (CHI)
  χ2is interested in (fo – fe)2/fe summed over all table entries: is the
observed number what you’d expect given the marginals?

2 2 2 2

χ (j,a) =∑(O − E) /E = (2 −.25) /.25+(3− 4.75) /4.75

+(500 −502)2 /502+(9500 −9498)2 /9498 =12.9 (p < .001)
  The null hypothesis (the two variables are independent) is
rejected with confidence .999,
  since 12.9 > 10.83 (the critical value for .999 confidence for a 1
degree of freedom χ2 distribution ).

Term = jaguar Term ≠ jaguar

expected: fe

Class = auto 2 (0.25) 500 (502)

observed: fo
Class ≠ auto 3 (4.75) 9500 (9498)

45
 Expected value, for instance in the up-left cell is:
P(c=auto)*P(T=jaguar)* #of cases = 502/10005 * 5/10005 *
10005 = 0.2508
χ2 statistic (CHI)

There is a “simpler” formula for 2x2 χ2:

A = #(t,c) C = #(¬t,c)

B = #(t,¬c) D = #(¬t, ¬c)

N=A+B+C+D
46
Feature selection via Mutual
Information

  Intraining set, choose k words which give most info

on the knowledge of the categories   The Mutual
Information between a word, class is:

p(U =ew,C=ec)
I(U,C)= ∑ ∑ p(U =ew,C=ec)log e ∈{0,1} e ∈{0,1} p(U
w c

=ew)p(C=ec)
  U=1 (U=0) means the document (does not) contains
w   C=1 (C=0) the document is (not) in class c   For

each word w and each category c   I(X,Y) = H(X) –

H(X|Y)
47
 =-Σi p(xi)logp(xi) + Σi p(yj) H(X|Y=yj)
  H is called the Entropy

Feature selection via MI (contd.)

  Foreach category we build a list of k most

discriminating terms
  For example (on 20 Newsgroups):
  sci.electronics: circuit, voltage, amp, ground,
copy, battery, electronics, cooling, …
  rec.autos: car, cars, engine, ford, dealer,
mustang, oil, collision, autos, tires, toyota, …
  Greedy: does not account for
correlations between terms   Why?
48
Feature Selection

  Mutual Information   Clear information-

theoretic interpretation   May select very
slightly informative frequent
terms that are not very useful for classification
  Chi-square  
Statistical foundation
  May select rare statistically correlated but
uninformative terms
  Just use the commonest terms?
  No particular foundation   In
practice, this is often 90% as good.

49
Example

50
Feature selection for NB

  In
general feature selection is necessary for
multivariate Bernoulli NB
  Otherwise you suffer from noise, multi-counting

  “Feature selection” really means something

different for multinomial NB - it means

dictionary truncation
  The multinomial NB model only has 1 feature
  This
“feature selection” normally isn’t needed
for multinomial NB, but may help a fraction with
quantities that are badly estimated.
51
Evaluating Categorization

  Evaluation must be done on test data that are

independent of the training data (usually a disjoint
set of instances).
  Classification
accuracy: c/n where n is the total
number of test instances and c is the number of test
instances correctly classified by the system
  Adequateif one class per document (and positive
and negative examples have similar cardinalities)
  Otherwise F measure for each class
  Resultscan vary based on sampling error due to
different training and test sets
  Average results over multiple training and test sets
(splits of the overall data) for the best results.

52
WebKB Experiment (1998)

  Classify webpages from CS departments into:

  student, faculty, course, project
  Train on ~5,000 hand-labeled web pages

53
  Results:

Student Faculty Person Project Course Departmt
Extracted 180 66 246 99 28 1
Correct 130 28 194 72 25 1
Accuracy: 72% 42% 79% 73% 89% 100%

Actually this is not accuracy but …

  Cornell, Washington, U.Texas, Wisconsin

  Crawl and classify a new site (CMU)

54
Most relevant features: MI

55
Naïve Bayes on spam email

56
Naïve Bayes Posterior Probabilities

  Classification results of naïve Bayes (the

class
with maximum posterior probability) are
usually fairly accurate
  However, due to the inadequacy of the
conditional independence assumption, the
actual posterior-probability numerical
estimates are not
  Output probabilities are commonly very close
to 0 or 1
  Correct estimation ⇒ accurate prediction, but

57
 correct probability estimation is NOT necessary
for accurate prediction (just need right ordering
of probabilities).
Naive Bayes is Not So Naive
  Naïve
Bayes: First and Second place in KDD-CUP 97 competition,
among 16 (then) state of the art algorithms
Goal: Financial services industry direct mail response prediction model:
Predict if the recipient of mail will actually respond to the advertisement
– 750,000 records.
  Robust to Irrelevant Features
Irrelevant Features cancel each other without affecting results
Instead Decision Trees can heavily suffer from this.
  Very good in domains with many equally important features
Decision Trees suffer from fragmentation in such cases – especially if
little data
  Agood dependable baseline for text classification (but not the
best)!
  Optimal if the Independence Assumptions hold: If assumed

58
 independence is correct, then it is the Bayes Optimal Classifier for
problem
  Very Fast: Learning with one pass of counting over the data; testing
linear in the number of attributes, and document collection size
  Low Storage requirements
Resources

  IIR 13
  Tom Mitchell, Machine Learning. McGraw-
Hill, 1997.   Clear simple explanation of
Naïve Bayes

59