15-780: Graduate Artificial

Intelligence

Inference in Bayesian networks

 Bayesian networks: Notations

Conditional P(Lo) = 0.5

probability tables Conditional
(CPTs) dependency
Le

P(Li | Lo) = 0.4 P(S | Lo) = 0.6

Li S
P(Li | ¬Lo) = 0.7 P(S | ¬Lo) = 0.2

Random variables
 A example problem
• An alarm system
B E
B – Did a burglary occur?
E – Did an earthquake occur?
A – Did the alarm sound off?
M – Mary calls A

J – John calls

J M
 Constructing a Bayesian network:
Revisited
• Step 1: Identify the random variables
• Step 2: Determine the conditional dependencies
- Select on ordering of the variables
- Add them one at a time
- For each new variable X added select the minimal subset of nodes
as parents such that X is independent from all other nodes in the
current network given its parents.
• Step 3: Populate the CPTs
- We will discuss this when we talk about density estimations
 Reconstructing a network
Suppose we wanted to add
a new variable to the
network: B E
R – Did the radio announce
that there was an
earthquake?
How should we insert it? A
R

J M
 Bayesian networks: Restrictions
and joint distributions
• Bayesian networks are directed acyclic graphs
(DAGs) x1 x2
- Otherwise a node will impact (indirectly) its
own probability making inference hard

x3 x4

x6 x5

This is NOT a valid Bayesian

network!
 Bayesian networks: Restrictions
and joint distributions
• Bayesian networks are directed acyclic graphs
(DAGs) x1 x2
- Otherwise a node will impact (indirectly) its
own probability making inference hard
• Given a Bayesian network the joint probability
distribution can be factored as: x3 x4

P ( X ) = ∏ p ( xi | Pa ( xi ))
i

where X is a vector of observations and Pa(xi) is x6 x5

the set of parent nodes of xi
 Using Bayesian networks
• Inference
- Computing joint distributions
- Inferring values of unobserved variables
• Structure learning
 Bayesian network: Inference
• Once the network is constructed, we
can use algorithms for inferring the B E
values of unobserved variables.
• For example, in our previous network P(B=1 | J, ¬M)?
the only observed variables are the
A
phone calls. However, what we are
really interested in is whether there
was a burglary or not.
• How can we determine that? J M
 Inference
• Lets start with a simpler question
- How can we compute a joint distribution from the
network?
- For example, P(B,¬E,A,J, ¬M)?
• Answer:
- That’s easy, lets use the network
 Computing: P(B,¬E,A,J, ¬M)
P(B,¬E,A,J, ¬M) = P(B)=.05
P(E)=.1
P(B)P(¬E)P(A | B, ¬E) B E
P(J | A)P(¬M | A)
= 0.05*0.9*.85*.7*.2
P(A|B,E) )=.95
= 0.005355 P(A|B,¬E) = .85
P(A| ¬ B,E) )=.5 A
P(A| ¬ B, ¬ E) = .05

P(J|A) )=.7 J M
P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Computing: P(B,¬E,A,J, ¬M)
P(B,¬E,A,J, ¬M) = P(B)=.05
P(E)=.1
P(B)P(¬E)P(A | B, ¬E) B E
P(J | A)P(¬M | A)
= 0.05*0.9*.85*.7*.2
P(A|B,E)
We can easily )=.95
compute a
= 0.005355 P(A|B,¬E) = .85
complete joint distribution.
P(A| ¬ B,E) )=.5 A
What about partial
P(A| ¬ B, ¬ E) = .05
distributions? Conditional
distributions?
P(J|A) )=.7 J M
P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Inference
• We are interested in queries of the form:
P(B | J,¬M)
• This can also be written as a joint: B E
P ( B , J , ¬M )
P ( B | J , ¬M ) =
P ( J , ¬M , B ) + P ( J , ¬M , ¬B )
A
chain rule

• How do we compute the new joint?

J M
 Computing partial joints
P ( B , J , ¬M )
P ( B | J , ¬M ) =
P ( B , J , ¬ M ) + P ( ¬B , J , ¬ M )

Sum all instances with these settings (the

sum is over the possible assignments to the
other two variables, E and A)
 Computing: P(B,J, ¬M)
P(B,J, ¬M) = P(B)=.05
P(E)=.1
P(B,J, ¬M,A,E)+ B E
P(B,J, ¬M, ¬ A,E) +
P(B,J, ¬M,A, ¬ E) +
P(B,J, ¬M, ¬ A, ¬ E) = P(A|B,E) )=.95
P(A|B,¬E) = .85
0.0007+0.00001+0.005+0. P(A| ¬ B,E) )=.5 A
0003 = 0.00601 P(A| ¬ B, ¬ E) = .05

P(J|A) )=.7 J M
P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Computing partial joints
P ( B , J , ¬M )
P ( B | J , ¬M ) =
P ( B , J , ¬ M ) + P ( ¬B , J , ¬ M )

Done! Sum all

instances with
these settings
 Computing: P(¬ B,J, ¬M)
P(¬ B,J, ¬M) = P(B)=.05
P(E)=.1
P(¬ B,J, ¬M,A,E)+ B E
P(¬ B,J, ¬M, ¬ A,E) +
P(¬ B,J, ¬M,A, ¬ E) +
P(¬ B,J, ¬M, ¬ A, ¬ E) = P(A|B,E) )=.95
P(A|B,¬E) = .85
0.00665+0.002+0.006+0.0 P(A| ¬ B,E) )=.5 A
345 = 0.049 P(A| ¬ B, ¬ E) = .05

P(J|A) )=.7 J M
P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Computing partial joints

P ( B, J , ¬M )
P ( B | J , ¬M ) =
P ( B, J , ¬M ) + P (¬B, J , ¬M )
0.006
= = 0.11
0.006 + 0.049
 Computing partial joints
P ( B , J , ¬M )
P ( B | J , ¬M ) =
P ( B , J , ¬ M ) + P ( ¬B , J , ¬ M )

Sum all instances with these settings (the

sum is over the possible assignments to the
other two variables, E and A)

But the number of possible assignments is exponential in

the unobserved variables?
That is, unfortunately, the best we can do. General querying
of Bayesian networks is NP-complete
 Inference in Bayesian networks if
NP complete (sketch)
• Reduction from 3SAT
• Recall: 3SAT, find satisfying assignments to the
following problem: (a ∨ b ∨ c) ∧ (d ∨ ¬ b ∨ ¬ c) …

What is P(Y)?
P(xi=1) = 0.5

P(xi=1) = (x1 ∨ x2 ∨ x3)

P(Y=1) = (x1 ∧ x2 ∧ x3 ∧ x4)

Y
 Other inference methods
• Convert network to a polytree
B E
- In a polytree no two nodes have
more than one path between them A
- For such a graph there is a linear
time algorithm J M
- However, converting into a polytree
requires a large increase in the size of
the graph (number of nodes)
B E

J M
 Why is inference in polytrees easy?
• In polytrees, given a variable X we
can always divide the other
variables into two sets:
E+: Variables ‘above’ X y1 y2

E-: Variables ‘below’ X X

• These sets are mutually exclusive
(why?) z2 z1
• Using these sets we can
efficiently compute conditional
and joint distributions
 Stochastic inference
• We can easily sample the joint
distribution to obtain possible
instances P(E)=.1
P(B)=.05
1. Sample the free variable B E
2. For every other variable:
- If all parents have been sampled,
sample based on conditional P(A|B,E) )=.95
P(A|B,¬E) = .85
distribution A
P(A| ¬ B,E) )=.5
P(A| ¬ B, ¬ E) = .05
We end up with a new set of
assignments for B,E,A,J and M
which are a random sample from P(J|A) )=.7 J M
the joint P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Stochastic inference
• We can easily sample the joint
distribution to obtain possible
instances P(E)=.1
P(B)=.05
1. Sample the free variable B E
2. For every other variable:
- If all parents have been sampled,
sample based on conditional P(A|B,E) )=.95
P(A|B,¬E) = .85
distribution A
P(A| ¬ B,E) )=.5
P(A| ¬ B, ¬ E) = .05
Is it always possible to
carry out this sampling
P(J|A) )=.7
procedure? why? J M
P(J|¬A) = .05
P(M|A) )=.8
P(M|¬A) = .15
 Using sampling for inference
• Lets revisit our problem: Compute P(B | J,¬M)
• Looking at the samples we can cound:
- N: total number of samples
- Nc : total number of samples in which the condition holds (J,¬M)
- NB: total number of samples where the joint is true (B,J,¬M)
• For a large enough N
- Nc / N ≈ P(J,¬M)
- NB / N ≈ P(B,J,¬M)
• And so, we can set
P(B | J,¬M) = P(B,J,¬M) / P(J,¬M) ≈ NB / Nc
 Using sampling for inference
• Lets revisit our problem: Compute P(B | J,¬M)
• Looking at the samples we can cound:
Problem: What if the condition rarely
- N: total number of samples
happens?
- Nc : total number of samples in which the condition holds (J,¬M)
- NB: total number ofWe wouldwhere
samples needthelots
jointand lots(B,J,¬M)
is true of
• For a large enoughsamples,
N and most would be wasted
- Nc / N ≈ P(J,¬M)
- NB / N ≈ P(B,J,¬M)
• And so, we can set
P(B | J,¬M) = P(B,J,¬M) / P(J,¬M) ≈ NB / Nc
 Weighted sampling
• Compute P(B | J,¬M)
• We can manually set the value of J to
1 and M to 0
B E
• This way, all samples will contain the
correct values for the conditional
variables
• Problems? A

J M
 Weighted sampling
• Compute P(B | J,¬M)
• Given an assignment to parents, we
assign a value of 1 to J and 0 to M.
B E
• We record the probability of this
assignment (w = p1p2) and we weight
the new joint sample by w
A

J M
 Weighted sampling algorithm for
computing P(B | J,¬M)
• Set NB,Nc = 0
• Sample the joint setting the values for J and M,
compute the weight, w, of this sample
• Nc = Nc+w
• If B = 1, NB = NB+w

• After many iterations, set

P(B | J,¬M) = NB / Nc
Bayesian networks for cancer
detection
 Constructing networks
• So far we assumed that the network is derived from
domain knowledge.
• That’s not always easy to do
• Examples:
- How are different regions in the brain related?
- How are terrorists related (social networks)?
 Inferring structure from data
• It is possible to infer structure if enough data is provided
• The goal would be to find a structure that leads to a
maximal likelihood*
MaxSP(D | S)
• Problems?

*More on this next week

 Inferring structure using maximum
likelihood principle
• The more edges we have, the higher
the likelihood! B E

A
P(M | A, E) ≥ P(M | A)

J M
Why?
- If the two are independent and we have
B E
perfect data, trivially holds
A
- We have more parameters to fit. If there is
some noise in the measurements, we
J M
would likely overfit the data
 Inferring structure using maximum
likelihood principle
• The more edges we have, the higher
the likelihood! B E

A
P(M | A, E) ≥ P(M | A)

J M
Solutions:
- Statistical tests
B E
- Penalty functions
A

J M
 Likelihood ratio test
• Given two competing models we can compute their
likelihood ratio

P( D | B) Model A
T ( D) = 2 log
P( D | A) X Y

Z
Always ≥ 0, but by how much?

Model B

X Y

Z
 Likelihood ratio test
• Given two competing models we can compute their
likelihood ratio

P( D | B) Model A
T ( D) = 2 log ~ χ2
P( D | A) X Y

Z
Always ≥ 0, but by how much?
The result is distributed according to χ2, which
is a distribution defined by the number of free Model B
parameters (the difference in complexity of the
two models) X Y

Z
 Likelihood ratio test
• Given two competing models we can compute their
likelihood ratio

P( D | B) Model A
T ( D) = 2 log ~ χ2
P( D | A) X Y

Z
Reject the more complicated model, unless the
ratio is high enough (can use, for example, the
Matlab function CHI2PDF to compute the
Model B
probability of seeing this ratio as a result of noise).

X Y

Z
 Penalty functions
• Likelihood ratio tests are appropriate for relatively small
problems (few variables)
• For larger problems we usually use a penalty function
• This function penalizes the likelihood based on the
complexity of the model
L(D | M) = P(D | M)-f(n)
where n is related to the number of parameters
• Most commonly used penalty function:
- AIC: Akaike's Information Criterion
- BIC: Bayesian information criterion
Structure learning for biology
 Important points
• Bayes rule
• Joint distribution, independence, conditional
independence
• Definition Bayesian networks
• Inference in Bayesian networks
• Constructing a Bayesian network