Title: Bayesian Networks: Review Representation, independence, and inference a few problems for your enjoym
1Bayesian Networks ReviewRepresentation,
independence, and inference( a few problems for
your enjoyment)
- Stanislav Funiak
- 10-701 Recitation, 3/30/2006
2Conference Submission Network
Done
Beer
in Time
Quality
Sleep1
Sleep2
Comm- ents1
Comm-ents2
Recom-mended
Accepted
3Bayesian Network
- DAG, each node is a variable
- For each var. X, CPD p(X Pa X)
- represents the distribution as
A,B,D,T,Q,S1,S2,C1,C2,R
4Smaller Network
C2t
Qt
C2f
Qf
p(S1f, Qt, C1f, C2t, Rt, At)
5Independence Relations
- where the force lies
- example independencies
-
-
6Factorization gt Independence relations
- We have seen
- starting from p factorizes according to G
- show p satisfies some independence relations
- Which independence assumptions in G?
p factorizes according to G
p satisfies indep. relations in G
7Independence relations encoded in G
- 1. Local Markov Assumptions
- X indep. NonDescendants(X) Pa X
8Independence relations encoded in G
- 2. absence of active trails
- Variables X indep of variables Ygiven Z if no
active trail betweenX and Y given Z
9Active trail Review
- A path X1 X2 Xk is an active trail
when variables OµX1,,Xn are observed if for
each consecutive triplet in the trail - Xi-1?Xi?Xi1, and Xi is not observed (Xi?O)
- Xi-1?Xi?Xi1, and Xi is not observed (Xi?O)
- Xi-1?Xi?Xi1, and Xi is not observed (Xi?O)
- Xi-1?Xi?Xi1, and Xi is observed (Xi2O), or one
of its descendents
10Independence relations encoded in G
- 2. absence of active trails
- Variables X indep of variables Ygiven Z if no
active trail betweenX and Y given Z
11Independence relations encoded in G
- 1. Local Markov Assumptions
- 2. absence of active trails (d-separation)
12Independence relations gt Factorization
p factorizes according to G
p satisfies indep. relations in G
p factorizes according to G
p satisfies indep. relations in G
13Independence relations gt Factorization
- Suppose
- Prove that p factorizes according to G
14Queries
- Marginal probability
- Most probable explanation (MPE)
- Active data collection
Variable elimination
15Conditioning on evidence
Q
S1
C1
C2
R
C2t
Qt
A
C2f
Qf
16Marginal probability query
Q
S1
C1
C2
R
A
17Most probable explanation
Q
S1
C1
C2
R
A
18Do we need this algorithm?
- Couldnt we just take argmax of marginals?
Take
19What you need to know
- Representation
- Independence relations
- local Markov assumption
- active trails / d-separation
- independence relations gt factorization
- Variable elimination
- marginal queries
- argmax queries