(** Graphical Models Certification Training: https://www.edureka.co/graphical-modelling-course **) This Edureka "Graphical Models" PPT answers the question "Why do we need Probabilistic Graphical Models?" and how are they compare to Neural Networks. It takes you through the basics of PGMs and gives real-world examples of its applications. Why do you need PGMs? What is a PGM? Bayesian Networks Markov Random Fields Use Cases Bayesian Networks & Markov Random Fields PGMs & Neural Networks Follow us to never miss an update in the future. Instagram: https://www.instagram.com/edureka_learning/ Facebook: https://www.facebook.com/edurekaIN/ Twitter: https://twitter.com/edurekain LinkedIn: https://www.linkedin.com/company/edureka

2. Agenda
Why do you need PGMs?
What is a PGM?
Bayesian Networks
Markov’s Random Fields
Use-Cases
Belief Networks in MRF
PGMs & Neural Networks
01
02
03
04
05
06
07

3. Why do you need PGMs?

4. Why do you need
Probabilistic Graphical
Models?
Probabilistic Graphical Models are rich frameworks for encoding
probability distributions over complex domains.
Compact Graphical Representation
PGM are frameworks used to create and
represent compact graphical models of complex
real world scenarios
01
Intuitive Diagrams of Complex Relationships
PGMs give us intuitive diagrams of complex
relationships between stochastic variables.
02
Convenient from Computational Aspect
PGMs are also convenient from computational
point of view, since we already have algorithms
for working with graphs and statistics.
03
Dynamic Simulation of Models
Using PGM we can simulate dynamics of
industrial establishments, create models, and
many other things.
04

5. What is a PGM?

6. What is a Probabilistic
Graphical Model?
WC
G
P
Consider you have 4 binary(Yes/No) variables.
Spot in the World Cup(Yes/No)
Performance in the Pre WC Tour (Yes/No)
Good Genetics(Yes/No)
F Good Form(Yes/No)

7. What is a Probabilistic
Graphical Model?
Consider you have 4 binary(Yes/No) variables.
WC
G
P
F

8. What is a Probabilistic
Graphical Model?
Components of a Graphical Model
Nodes = Random Variables
WC
G
P
F

9. What is a Probabilistic
Graphical Model?
Components of a Graphical Model
Edges = Inter-Nodal Dependencies

10. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P
M
G
Probabilistic
Graphical
Models

11. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P Probabilistic
G M
The nature of the problem that we are generally interested to
solve or the type of queries we want to make are all
probabilistic because of uncertainty. There are many reasons
that contributes to it.

12. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P
M
G
Probabilistic
Graphical
Models

13. What is a Probabilistic
Graphical Model?
So, What is a PGM?
G Graphical
P M
Graphical representation helps us to visualise better and
So, we use Graph Theory to reduce the no of relevant
combinations of all the participating variables to represent
the high dimensional probability distribution model more
compactly.

14. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P
M
G
Probabilistic
Graphical
Models

15. What is a Probabilistic
Graphical Model?
So, What is a PGM?
M Models
P G
A Model is a declarative representation of a real world
scenario or a problem that we want to analysis. It is
represented by using any mathematical tools like graph or
even simply by an equation.

16. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P
M
G
Probabilistic
Graphical
Models

17. What is a Probabilistic
Graphical Model?
So, What is a PGM?
P G M
Probabilistic Graphical Models (PGM) is a technique of
compactly representing a joint distribution by exploiting
dependencies between the random variables. It also allows us
to do inference on joint distributions in a computationally
cheaper way than the traditional methods.

18. What is a Probabilistic
Graphical Model?
Probability
A
A
A
A
A
B
B
B
B
B
B
C
C
C
C
C
C
What is the Probability of A?
Solution:
• Count all As and Divide it by Total number of Possibilities.
• P(A) = (#A)/(#A+#B+#C)

19. What is a Probabilistic
Graphical Model?
Conditional Probability
A
A
A A
B
B
B
B
B
What is the Probability of A&B?
Solution:
• B should occur when A is already happening.
• P(A&B) = P(A) * P(B|A) or P(B) * P(A|B).

20. What is a Probabilistic
Graphical Model?
Joint, Probability and Marginal Distributions
• The Joint Probability Distribution describes how two or more
variables are distributed simultaneously. To get a probability from
the joint distribution of A and B, you would consider P(A=a and
B=b).
• The Conditional Probability Distribution looks at how the
probabilities of A are distributed, given a certain value, say, for B,
P(A=a| B=b).
• The Marginal Probability Distribution is one that results from
taking mean over one variable to get the probability distribution of
the other.
For Example, the marginal probability distribution of A when A & B
are related would be given by the following;
‫׬‬𝑩
P(a|b) P(b)db

21. Bayesian Networks

22. Bayesian Networks
Bayesian Probability
A
A
A A
B
B
B
B
B
P(One Event | Another Event)
We have seen earlier:
• P(A&B) = P(A) * P(B|A) or P(B) * P(A|B)
From here, we could isolate either P(B|A) or P(A|B) and
compute from simpler probabilities.

23. Bayesian Networks
Bayes Theorem
A
A
A A
B
B
B
B
B
We have seen earlier:
• P(A|B) = P(B|A) * P(A)/P(B)
• P(B|A) = P(A|B) * P(B)/P(A)

24. Bayesian Networks
Bayes Network
A Bayes Network is a structure that can be represented as a
Direct Acyclic Graph.
1. It allows a compact representation of the distribution from
the chain rule of Bayes Networks.
2. It observes conditional independence relationships
between random variables.
WC
G
P
F
A DAG(Direct Acyclic Graph) is a finite
directed graph with no directed cycles.

25. Bayesian Networks
Bayes Theorem : Example
WC
G
P
F
Genes P(Genes)
Good 0.2
Bad 0.8
In Form P(Form)
Yes 0.7
No 0.3
Condition No
Spot
Spot
given
Bad
Performance
0.95 0.05
Okay
Performance
0.8 0.2
Brilliant
Performance
0.5 0.5
Condition Bad Okay Brilliant
Good Genes,
Good Form
0.5 0.3 0.2
Good Genes,
Bad Form
0.8 0.15 0.05
Bad Genes,
Good Form
0.8 0,1 0.1
Bad Genes,
Bad Form
0.9 0.08 0.02

26. Bayesian Networks
Bayes Theorem : Example
WC
G
P
F
What should you think about?
• Does a spot in the WC team depend on
Genetics?
• Does a spot in the WC team depend on
Genetics if you know someone is in good
form?
• Does a spot in the WC team depend on
Genetics if you know the performance in
the Pre WC Tour?

27. Bayesian Networks
Bayes Theorem : Example
WC
G
P
F
How this works.
• Each node in the Bayes Network will
have a CPD associated with it.
• If the node has parents, the associated
CPD represents P(value| parent’s value)
• If a node has no parents, the CPD
represents P(value), the unconditional
probability of the value.

28. Markov’s Random Fields

29. Markov’s Random Fields
Undirected Graphical Models
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
Probability Distribution of the variables in the graph can
factorised as individual clique potential functions.

30. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

31. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

32. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

33. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

34. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C) ϕ(B,D) ϕ(C,E) ϕ(D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

35. Markov’s Random Fields
Cliques
B CA
ED
• P(A,B,C,D,E) ∝ ϕ(A,B) ϕ(B,C,D) ϕ(C,D,E)
P(X) =
1
𝑍
ς 𝑐∈𝑐𝑙𝑖𝑞𝑢𝑒𝑠(𝐺) ϕ 𝑐 (𝑥 𝑐 ) potential functions

36. Markov’s Random Fields
Markov Random Fields
B CA
ED
• Paths between A and C ;
A-B-C
A-B-D-E-C

37. Markov’s Random Fields
Markov Random Fields
B CA
ED
• Paths between A and C ;
A-B-C
A-B-D-E-C

38. Markov’s Random Fields
Markov Random Fields
B CA
ED
• Any two subsets of variables are conditionally independent,
given a separating subset.
• {B,D},{B,E} & {B,D,E} are the separating subsets.

39. Use- Cases

40. Use Cases
Applications of PGMs
Google is based on a very
simple graph algorithm
called page rank.
Netflix, Amazon, Facebook
all use PGMs to
recommend what is best for
you.

41. Use Cases
PGMs can also apparently
infer whether one is a Modi
Supporter or Kejriwal
Supporter.
FiveThirtyEight is a company
that makes predictions about
American Presidential Polls using
PGMs.
Applications of PGMs

42. Bayesian Networks &
Markov Random Fields

43. Belief Networks &
Markov Random Fields
Bayes Nets as MRFs
BA
BA
Bayes Network
MRF
P(A,B) = P(A) * P(B|A)
P(A,B) ∝ ϕ(A,B)

44. Belief Networks &
Markov Random Fields
Bayes Nets as MRFs
Bayes Network
MRF
P(A,B) = P(A)P(B|A)P(C|B)
P(A,B) ∝ ϕ(A,B) ϕ(B,C)
BA C
BA C

45. Belief Networks &
Markov Random Fields
Bayes Nets as MRFs : Chains
Bayes Network
MRF
P(A,B,C) = P(A)P(B|A)P(C|B)
P(A,B,C) ∝ ϕ(A,B) ϕ(B,C)
ϕ(A,B) ← P(A)P(B|A)
ϕ(B,C) ← P(C|B)
Parameterization is not unique
BA C
BA C

46. Belief Networks &
Markov Random Fields
Bayes Nets as MRFs : Shared Parents
Bayes Network
MRF
P(A,B,C) = P(A)P(B|A)P(C|B)
P(A,B,C) ∝ ϕ(A,B) ϕ(A,C)
ϕ(A,B) ← P(A)P(B|A)
ϕ(A,C) ← P(C|A)
B
A
C
B
A
C

47. Belief Networks &
Markov Random Fields
Bayes Nets as MRFs : Shared Child
Bayes Network
MRF
P(A,B,C) = P(A)P(B)P(C|A,B)
A and B are dependent given C
P(A,B,C) ∝ ϕ(A,C) ϕ(B,C)
A and B are independent given C
C
A B
C
A B

48. Belief Networks &
Markov Random Fields
Converting Bayes Nets to MRFs : Moralizing Parents
P(A,B,C) ∝ ϕ(A,C) ϕ(B,C)
A and B are independent given C
C
A B
• Moralize all co-parents.
• Lose marginal independence of parents
undirecteddirected

49. PGMs and Neural Networks

50. PGMs & Neural
Networks
The Boyfriend Problem

51. PGMs & Neural
Networks
The Boyfriend Problem

52. PGMs & Neural
Networks
The Boyfriend Problem

53. PGMs & Neural
Networks
The Boyfriend Problem
• N common friends
• N bit vector = {0,1,0,1,…} ; 1 is a boost, 0 means no contact
• Objective : find the set of friends that I should ask to boost me,
i.e. the best vector.

54. PGMs & Neural
Networks
Solution 1 : Neural Networks
Yes/No
More Friends/
Happening Social
Life(M)
Me
Friends(N)

55. PGMs & Neural
Networks
Solution 1 : Neural Networks
More Friends/
Happening Social
Life(M)
Yes/No
Friends(N)
Me

56. PGMs & Neural
Networks
Solution 1 : Neural Networks
But, Why?
Me

57. PGMs & Neural
Networks
Solution 2 : Probabilistic Graphical Models
Approval
from N
Friends/
Observation
from N
nodes
Me
0
1
0
1
0
0
1
1
Date Acceptance/
Observable Event
Impression/
Random Variable

58. PGMs & Neural
Networks
Solution 2 : Probabilistic Graphical Models
Date Acceptance/
Observable Event
Approval
from N
Friends/
Observation
from N nodes
Me
Impression/ Random
Variable
P(impression| vector of all approvals) ← P(approval from friend i| impression)
P(all approvals| impression) = product of all P(approval i | impression)