This document provides an overview of text classification and the Naive Bayes machine learning algorithm. It defines text classification as assigning categories or labels to documents, and discusses different approaches like human labeling, rule-based classification, and machine learning. Naive Bayes is introduced as a simple supervised learning method that calculates the probability of documents belonging to different categories based on word frequencies. The document then reviews probability concepts and shows how Naive Bayes makes the "naive" assumption that words are conditionally independent given the topic to classify documents probabilistically using Bayes' theorem.

1. Text Classification and Naïve Bayes
An example of text classification
Definition of a machine learning problem
A refresher on probability
The Naive Bayes classifier
1

2. Google News
2

3. Different ways for classification
Human labor (people assign categories to every incoming
article)
Hand-crafted rules for automatic classification
 If article contains: stock, Dow, share, Nasdaq, etc.  Business
 If article contains: set, breakpoint, player, Federer, etc.  Tennis
Machine learning algorithms
3

4. What is Machine Learning?
4
Definition: A computer program is said to learn from
experience E when its performance P at a task T
improves with experience E.
Tom Mitchell, Machine Learning, 1997
Examples:
- Learning to recognize spoken words
- Learning to drive a vehicle
- Learning to play backgammon

5. Components of a ML System (1)
Experience (a set of examples that combines together
input and output for a task)
 Text categorization: document + category
 Speech recognition: spoken text + written text
Experience is referred to as Training Data. When training
data is available, we talk of Supervised Learning.
Performance metrics
 Error or accuracy in the Test Data
 Test Data are not present in the Training Data
 When there are few training data, methods like ‘leave-one-out’ or
‘ten-fold cross validation’ are used to measure error.
5

6. Components of a ML System (2)
Type of knowledge to be learned (known as the target
function, that will map between input and output)
Representation of the target function
 Decision trees
 Neural networks
 Linear functions
The learning algorithm
 C4.5 (learns decision trees)
 Gradient descent (learns a neural network)
 Linear programming (learns linear functions)
6
Task

7. Defining Text Classification
7
XdX∈d
},,,{ 21 Jccc =C
D cd,
C×∈Xcd,
C→X:γ
γ=Γ D)(
the document in the multi-dimensional space
a set of classes (categories, or labels)
the training set of labeled documents
Target function:
Learning algorithm:
=cd, “Beijing joins the World Trade Organization”, China
cd =)(γ =)(dγ China

8. Naïve Bayes Learning
8
∏≤≤∈∈
==
dnk
k
CcCc
MAP ctPcPdcPc
1
)|(ˆ)(ˆmaxarg)|(ˆmaxarg
cd =)(γ
Learning Algorithm: Naïve Bayes
Target Function:
)|()(maxarg)|(maxarg cdPcPdcPc
CcCc
MAP
∈∈
==
)(cP
)|( cdP
The generative process:
)|( dcP
a priori probability, of choosing a category
the cond. prob. of generating d, given the fixed c
a posteriori probability that c generated d

9. A Refresher on Probability
9

10. Visualizing probability
A is a random variable that denotes an uncertain event
 Example: A = “I’ll get an A+ in the final exam”
P(A) is “the fraction of possible worlds where A is true”
10
Worlds in
which A
is true
Slide: Andrew W. Moore
Worlds in which A is false
Event space of all possible
worlds. Its area is 1.
P(A) = Area of the blue
circle.

11. Axioms and Theorems of Probability
Axioms:
 0 <= P(A) <= 1
 P(True) = 1
 P(False) = 0
 P(A or B) = P(A) + P(B) – P(A and B)
Theorems:
 P(not A) = P(~A) = 1 – P(A)
 P(A) = P(A ^ B) + P(A ^ ~B)
11

12. Conditional Probability
P(A|B) = the probability of A being true, given that we
know that B is true
12
F
H
H = “I have a headache”
F = “Coming down with flu”
P(H) = 1/10
P(F) = 1/40
P(H/F) = 1/2
Slide: Andrew W. Moore
Headaches are rare and flu
even rarer, but if you got that flu,
there is a 50-50 chance you’ll
have a headache.

13. Deriving the Bayes Rule
13
)(
)(
)|(
BP
BAP
BAP
∧
=Conditional Probability:
)()|()( BPBAPBAP =∧Chain rule:
)()|()()( APABPABPBAP =∧=∧
Bayes Rule:
)(
)()|(
)|(
AP
BPBAP
ABP =

14. Back to the Naïve Bayes Classifier
14

15. Deriving the Naïve Bayes
15
)(
)()|(
)|(
AP
BPBAP
ABP = (Bayes Rule)
21,cc 'dGiven two classes and the document
)'(
)|'()(
)'|( 11
1
dP
cdPcP
dcP =
)'(
)|'()(
)'|( 22
2
dP
cdPcP
dcP =
We are looking for a that maximizes the a-posterioriic )'|( dcP i
)'(dP (the denominator) is the same in both cases
)|()(maxarg cdPcPc
Cc
MAP
∈
=Thus:

16. Estimating parameters for the
target function
We are looking for the estimates and
16
)(ˆ cP )|(ˆ cdP
P(c) is the fraction of possible worlds where c is true.
N
N
cP c
=)(ˆ N – number of all documents
Nc – number of documents in class c
d is a vector in the space X
)|,,,()|( 2 ctttPcdP dni =
where each dimension is a term:
)()|()( BPBAPBAP =∧By using the chain rule: we have:
(P
),,...,(),,...,|()|,,,( 2212 cttPctttPctttP ddd nnni =
...=

17. Naïve assumptions of independence
1. All attribute values are independent of each other given
the class. (conditional independence assumption)
2. The conditional probabilities for a term are the same
independent of position in the document.
We assume the document is a “bag-of-words”.
17
∏≤≤
==
d
d
nk
kni ctPctttPcdP
1
2 )|()|,,,()|( 
∏≤≤∈∈
==
dnk
k
CcCc
MAP ctPcPdcPc
1
)|(ˆ)(ˆmaxarg)|(ˆmaxarg
Finally, we get the target function of Slide 8:

18. Again about estimation
18
For each term, t, we need to estimate P(t|c)
∑ ∈
=
Vt ct
ct
T
T
ctP
' '
)|(ˆ
Because an estimate will be 0 if a term does not appear with a class
in the training data, we need smoothing:
||)(
1
)1(
1
)|(ˆ
' '' ' VT
T
T
T
ctP
Vt ct
ct
Vt ct
ct
∑∑ ∈∈
+
+
=
+
+
=Laplace
Smoothing
|V| is the number of terms in the vocabulary
Tct is the count of term t in all documents of class c

19. An Example of classification with
Naïve Bayes
19

20. Example 13.1 (Part 1)
20
Training
set
docID c = China?
1 Chinese Beijing Chinese Yes
2 Chinese Chinese Shangai Yes
3 Chinese Macao Yes
4 Tokyo Japan Chinese No
Test set 5 Chinese Chinese Chinese Tokyo Japan ?
Two classes: “China”, “not China”
N = 4 4/3)(ˆ =cP 4/1)(ˆ =cP
V = {Beijing, Chinese, Japan, Macao, Tokyo}

21. Example 13.1 (Part 1)
21
Training
set
docID c = China?
1 Chinese Beijing Chinese Yes
2 Chinese Chinese Shangai Yes
3 Chinese Macao Yes
4 Tokyo Japan Chinese No
Test set 5 Chinese Chinese Chinese Tokyo Japan ?
7/3)68/()15()|Chinese(ˆ =++=cP
14/1)68/()10()|Japan(ˆ)|Tokyo(ˆ =++== cPcP
9/2)63/()11()|Chinese(ˆ =++=cP
9/2)63/()11()|Japan(ˆ)|Tokyo(ˆ =++== cPcP
Estimation Classification
∏≤≤
∝
dnk
k ctPcPdcP
1
)|()()|(
0001.09/29/2)9/2(4/1)|(
0003.014/114/1)7/3(4/3)|(
3
5
3
5
≈⋅⋅⋅∝
≈⋅⋅⋅∝
dcP
dcP

22. Summary: Miscellanious
Naïve Bayes is linear in the time is takes to scan the data
When we have many terms, the product of probabilities
with cause a floating point underflow, therefore:
For a large training set, the vocabulary is large. It is better
to select only a subset of terms. For that is used “feature
selection” (Section 13.5).
22
∑≤≤∈
+=
dnk
k
Cc
MAP ctPcPc
1
)|(log)(ˆ[logmaxarg