Transcript Probability --- Part b

Discrete Math
CS 2800
Prof. Bart Selman
[email protected]
Module
Probability --- Part b)
Bayes’ Rule
Random Variables
1
Bayes’ Theorem
How to assess the probability that a particular event will occur on the
basis of partial evidence?
Examples:
What is the likelihood that people who test positive to a particular
disease
(e.g., HIV), actually have the disease?
What is the probability that an e-mail message is spam?
Key idea: one should factor in additional information regarding
occurrence of events.
2
Assume that with respect to events F and E (“E” for “Evidence”):
We know P(F) – probability that event F occurs
(e.g. probability that email message is spam;
this is given by what fraction of email is spam)
We also know event E has occurred.
(e.g., email message contains words “sale” and “bargain”)
Therefore the probability conditional probability that F occurs given
that E occurs, P(F|E), is a more realistic estimate that F occurs than P(F).
How do we compute P(F|E)?
E.g., based on P(F), P(E|F), and P(E| ¬F)
Note: ¬F is also referred to as complement of F (FC or F).
E
Evidence
Original
Belief
Hypothesis
(Prior Probability)
P(F)
Bayesian
Inference
Modified
Theory
Belief
P(F|E)
Box A
Box B
Experiment: Pick one box at random (p = 0.5) and than a ball at
random from that box.
Assume you picked a red ball.
What’s the probability that it came form the left box? > 0.5 ? Why?
Define:
E – you choose a red ball. (therefore ¬ E – you choose the green ball)
F – you choose the left box. (therefore ¬ F– you choose the right box)
We want to know P(F|E)
P(F|E)?
E – red color
F – left box
What we know:
P(E|F) = 7/9
and P(E|¬F) = 3/7
Given that the boxes are selected at random: P(F) = P(¬F)=1/2
P(F|E) = P(E∩F)/P(E)  so we need to compute P(E∩F) and P(E).
We know P(E|F) = P(E∩F)/P(F).
So, P(E∩F) = P(E|F) P(F) = 7/9 * 1/2 = 7/18.
What about P(E)? Note that P(E) = P(EF) +P (E∩ ¬F). Why?
Note also that P (E ∩ ¬F) = P(¬F) P(E|¬F) = 1/2 * 3/7 = 3/14
So, P(E) = P(EF) +P (E∩ ¬F) = 7/18 + 3/14 = 38/63
And therefore P(F|E) = P(E∩F)/P(E) = (7/18) / (38/63) = 49/76 0.645
Indeed > 0.5 phew!
Original Belief
there is a 0.5 that you will pick left box
(P(F)).
Bayesian
Inference
Modified Belief
Increased to 0.65 Probability
(P(F|E))
Concrete (new) Evidence
Red ball picked (E)
Theorem: Bayes’ Theorem
Suppose that E and F are events from a sample space S such
that P(E) ≠0 and P(F)≠0. Then
Proof:
P( F | E )  P( E  F ) / P( E )
P( E | F )  P( E  F ) / P( F )
P( E | F ) P( F )
P( F | E ) 
P( E | F ) P( F )  P( E | F C ) P( F C )
So, P( E  F )  P ( E | F ) P( F ) therefore
P( E | F ) P( F )
P( F | E ) 
P( E )
P( E | F ) P( F )
P( F | E ) 
P( E )
We only need an exp ression for P( E )
E  E  S  E  (F  F C )  (E  F )  (E  F C )
So
P( E )  P( E  F )  P( E  F C )  P( E | F ) P( F )  P( E | F C ) P( F C )
So
P( E | F ) P( F )
P( E | F ) P( F )
P( F | E ) 

P( E )
P( E | F ) P( F )  P( E | F C ) P( F C )
Example --- A Classic!
Suppose that 1 person in 100,000 has a particular rare disease. There is an
good test for the disease that is correct in 99% of the time when given
to someone with the disease; it is correct in 99.5% of the time when given
to someone without the disease.
Find:
a) Probability that someone who tests positive has the disease.
b) Probability that someone who tests negative does not have the disease.
1 person in 100,000 has rare disease. Test correct
in 99% of the time when given to someone
Solution:
with the disease; correct in 99.5% of the time
Always start by defining the events! when given to someone without the disease.
a) F – the person has the disease
E – the person tests positive to the disease
P(F|E) – probability of having the disease given positive test
P(F)=1/100,000 = 0.00001; P(FC) = 0.99999 Why these probabilities?
Most easily measured!
P(E|F) = 0.99; P(EC|F) = 0.01
P(E|FC) = 0.005
P( E | F ) P( F )
P( F | E ) 
P( E | F ) P( F )  P( E | F ) P( F )
(0.99)(0.00001)
 0.002
(0.99)(0.00001)  (0.005)(0.99999)
C
C

Only 0.2% of people who test positive actually have the disease!!!
Counter –intuitive. That’s why it’s a classic!
(note: test is good the disease is rare; but context matters.)
b) F – the person has the disease
E – the person tests positive to the disease
P(FC|EC) – probability of not having the disease given negative test
P(F) = 1/100,000 = 0.00001; P(FC) = 0.99999
P(E|F) = 0.99; P(EC|F) = 0.01
P(E|FC) = 0.005
P( E C | F C ) P( F C )
P( F | E ) 

C
C
C
C
P( E | F ) P( F )  P( E | F ) P( F )
(0.995)(0.999991)
 0.9999999
(0.995)(0.99999)  (0.01)(0.00001)
C
C
That’s… pretty good!
Generalized Bayes’ Theorem
Suppose that E is an event from a sample space S and F1, F2 ,…, Fn are
 ni1 Fi  S
mutually exclusive events such that
Asume that P(E) ≠ 0 and P(Fi) ≠ 0 for i=1, 2,…, n. Then
P( F j | E ) 
P( E | F j ) P( F j )
n
 P( E | Fi ) P( Fi )
i 1
P(E)
Compare:
P( F | E ) 
P( E | F ) P( F )
P( E | F ) P( F )  P( E | F C ) P( F C )
Bayesian Spam Filters
17 17
Applying Bayes’ Theorem:
SPAM or HAM?
Let our sample space or universe be the set of emails. (So, we’re sampling from
the space of possible emails.)
Let S be the event a message is spam; hence is the event a message is not spam
Let E be the event a message contains a word w.
Since we have no idea of likelihood of SPAM,
we assume P(S)=P(SC)=1/2.
How do we get p( E | S ) and
Can we do better?
p( E | S ) ?
18
Estimations
Note these are estimates based on frequencies in samples.
19 19
p( E | S )  p(w)
p( E | S )  q(w)
Estimation Continued
So,
becomes
So, what do we want for
p(w) and q(w) ??
What’s the max prob and
what’s the min? When
do we get 0.5?
So, a quite straightforward formula for our first
Bayesian spam filter!
Note P(S) = P(SC) = ½ divides
20
Spam based on single words?
Probabilities based on single words: Bad Idea
– False positives AND false negatives a plenty
Calculate based on n words, assuming each event Ei|S (Ei|SC) is
independent (not true but reasonable approximation)
P(S) = P(SC).
Derivation see Sect. 6.3.
21 21
Final Approximation
Compare to single word:
22 22
How do we use this?
User must train the filter based on messages in his/her inbox to estimate
probabilities.
The program or user must define a threshold probability r:
If
, the message is considered spam.
Gmail: Train on all users! (note: report spam button)
23 23
Example
Suppose the filter has the following data
Threshold Probability: .9
“Nigeria” occurs in 250 of 2000 spam messages
“Nigeria” occurs in only 5 of 1000 non-spam messages
Let’s try to estimate the probability, using the process we just defined
24 24
Example Cont.
Step 1: Find the probability that the message has the
word “Nigeria” in it and is spam.
– p(Nigeria) = 250 / 2000 = 0.125
Step 2: Find the probability that the message has the
word “Nigeria” in it and is not spam.
– q(Nigeria) = 5 / 1000 = 0.005
25 25
Example Cont.
Since we are assuming that it is equally likely that an
incoming message is or is not spam, we can
estimate the probability with this equation:
r(Nigeria) =
p(Nigeria)
p(Nigeria) + q(Nigeria)
26 26
Example Cont.
0.125____
0.125 + 0.005
= 0.125
0.130
= 0.962
Since r(Nigeria) is greater than the threshold of 0.9, we can reject
this message as spam.
27 27
Multiple Words
2000 Spam messages; 1000 real messages
“Nigeria” appears in 400 spam messages
“Nigeria” appears in 60 real messages
“bank” appears in 200 spam and 25 real messages
Threshold Probability: .9
Let’s calculate the probability that message with “Nigeria” and “bank” is
spam.
28 28
Example Cont.
Step 1: Find the probability that the message has the word “Nigeria” in it
and is spam.
– p(Nigeria) = 400 / 2000 = 0.2
Step 2: Find the probability that the message has the word “Nigeria” and is
not spam.
– q(Nigeria) = 60 / 1000 = 0.06
Step 3: Find the probability that the message contains the word “bank” and
is spam.
– p(bank) = 200 / 2000 = 0.1
Step 4: Find the probability that the message contains the word “bank” and
is not spam.
– q(bank) = 25 / 1000 = 0.025
29
Example Cont
Using our approximation, we have:
r(Nigeria,bank) =
p(Nigeria) * p(bank)
p(Nigeria) * p(bank) + q(Nigeria) * q(bank)
30 30
Example Cont.
Using our approximation, we have:
r(Nigeria,bank) =
p(Nigeria) * p(bank)
p(Nigeria) * p(bank) + q(Nigeria) * q(bank)
r(Nigeria,bank)
=
(0.2)(0.1)
(0.2)(0.1) + (0.6)(0.025)
= 0.930
This message will be rejected however since we set the threshold probability at 0.9.
Concludes Bayes Reasoning
31 31
Probability Paradox I
32
C
Magic Dice:
Or How to Win Every Time!
D
a) You select any one of the four dice (A, B, C, or D).
b) I’ll select another.
Both dice are thrown, highest number wins throw.
Do series of 10 throws. The person with the most
highest throws wins the series. (I.e. die “more likely to get a higher number” wins.)
B
A
Claim: In a game of 'The Best of Ten Throws’, I will almost certainly win
--- no matter which die you pick!!
Why is this strange?
Say, you pick die A. Let’s assume, die B is better. So, I pick B.
But, then, next game & next person picks B. Let’s assume C is better. I’ll select C.
Next person, will pick C. I’ll pick D.
But, could such a set of
Next person, will pick D… Hmm… ??
dice exist?
I’ll pick A and will win!!
A < B < C < D …. < A !! Failure of transitivity!
Surprisingly, yes!
Magic Dice
Prob(D wins over C) = 2/3
D 2/6 + (4/6)* (1/2) = 4/6
Prob(A wins over D)
= 4/6 = 2/3
A < B < C < D …. < A !!
A
C
Prob(C wins over B) = 2/3
Prob(B wins over A) = 4/6 = 2/3
B
since
3/6 + (3/6)* (2/6) = 4/6
(i.e. Prob(A wins over B) = 1/3)
How about expected value of dice throw?
Transitive or not?? E.g. E[B] = 16/6.
YES!!
E[B] < E[A] = E[C] < E[D]
16/6 < 18/6 = 18/6 < 20/6
http://www.sciencenews.org/20020420/mathtrek.asp
Random Variables and Distributions
35
Random Variables
For a given sample space S, a random variable (r.v.) is any real valued
function on S, i.e., a random variable is a function that assigns a real
number to each possible outcome
Numbers
Sample space
S
-2
0
Suppose our experiment is a roll of 2 dice. S is set of pairs.
Example random variables:
X = sum of two dice.
X((2,3)) = 5
Y = difference between two dice.
Y((2,3)) = 1
Z = max of two dice.
Z((2,3)) = 3
2
Random variable
Suppose a coin is flipped three times. Let X(t) be the random variable
that equals the number of heads that appear when t is the
outcome.
X(HHH) = 3
X(HHT) = X(HTH) = X(THH) = 2
X(TTH) = X(THT) = X(HTT) = 1
X(TTT) = 0
Note: we generally drop the argument! We’ll just say the
“random variable X” (even though it’s technically a function).
And write e.g. P(X = 2) for “the probability that the random variable X(t) takes
on the value 2”.
Or P(X=x) for “the probability that the random variable X(t) takes
on the value x.”
Distribution of Random Variable
Definition:
The distribution of a random variable X on a sample space S is the set of
pairs (r, p(X=r)) for all r  X(S), where p(X=r) is the probability that X
takes the value r.
A distribution is usually described specifying p(X=r) for each r  X(S).
A probability distribution on a r.v. X is just an allocation of
the total probability mass, 1, over the possible values of X.
38
The Birthday Paradox
41
A: 23
Birthdays
a) 23
How many people have to be in a room to assure that the probability
that at least two of them have the same birthday is greater than 1/2? b) 183
c) 365
d) 730
Let pn be the probability that no people share a birthday among n people
in a room.
Then 1 - pn is the probability that 2 or more share a birthday.
We want the smallest n so that 1 - pn > 1/2.
For L options
answer is in
the order
of sqrt(L) ?
Informally, why??
Hmm. Why does such an n exist? Upper-bound?
Birthdays
Assumption:
Birthdays of the people are independent.
Each birthday is equally likely and that there are 366 days/year
Let pn be the probability that no-one shares a birthday among n people in a room.
What is pn? (“brute force” is fine)
Assume that people come in certain order; the probability that the second person
has a birthday. Different than the first is 365/366; the probability that the third person
has a different birthday. Form the two previous ones is 364/366.. For the jth person
we have (366-(j-1))/366.
So,
365 364 363 367  n
pn 

366 366 366
366
365 364 363 367  n
1  pn  1 

366 366 366
366
After several tries, when n=22 1= pn = 0.475.
n=23 1-pn = 0.506
Relevant to “hashing”. Why?
Probability of a Collision in Hashing
Functions
A hashing function h(k) is a mapping of the
keys (or records, e.g., SSN, around
300x 106 in the US) to a much smaller
storage location. A good hashing fucntion
yields few collisions. What is the
probability that no two keys are mapped
to the same location by a hashing function?
Assume m is the number available storage
locations, so the probability
of mapping a key to a location is 1/m.
Assuming the keys are k1, k2, kn, the
probability of mapping the jth record to a
free location is after the first (j-1) records
is (m-(j-1))/m.
From Birthday Problem to
Hashing Functions
m 1 m  2 m  n 1

m
m
m
m 1 m  2 m  n 1
1  pn  1 

m
m
m
pn 
Given a certain m, find the smallest n
Such that the probability of a collision
is greater than a particular threshold p.
It can be shown that for p>1/2,
n 1.177
m
m = 10,000, gives n = 117. Not that many!
45