Bayesian networks in AI

AI, Uncertainty & Bayesian Networks
(Slide for an AI course lecture in SNU)

1. AI, Uncertainty & Bayesian Networks
2015-03-10 / 03-12
Kim, Byoung-Hee
BiointelligenceLaboratory
Seoul National University
http://bi.snu.ac.kr
Artificial Intelligence: Cognitive Agents

2. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 2
A Bayesian networkis a graphical model
for probabilistic relationships among a set of variables

3. ACM Turing Award
Nobel Prize in Computing
2011 Winner: Judea Pearl (UCLA)
For fundamental contributions to artificial intelligence
through the development of a calculus for probabilistic
and causal reasoning
Invention of Bayesian networks
Pearl's accomplishments have “redefined the term
'thinking machine’” over the past 30 years
BN mimics “the neural activities of the human brain,
constantly exchanging messages without benefit of a
supervisor”
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 3

4. Notes
Related chapters in the textbook (AIMA 3
rd
ed. by Russell and Norvig)
Ch. 13 Quantifying Uncertainty
Ch. 14 Probabilistic Reasoning (14.1~14.2)
Ch. 20 Learning Probabilistic Models (20.1~20.2)
On the reference ‘A Tutorial on Learning with Bayesian
Networks by David Heckerman’
It is very technical but covers insights and comprehensive
backgrounds on Bayesian networks
This lecture covers the ‘Introduction’ section
This lecture is an easier introductory tutorial
Both contents in the textbook and Heckerman’s tutorial is fairly
mathematical
This lecture covers basic concepts and tools to understand
Bayesian networks
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 4

5. Contents
Bayesian Networks: Introduction
Motivating example
Decomposing a joint distribution of variables
d-separation
A mini Turing test in causalconversation
Correlation & causation
AI & Uncertainty
Bayesian Networks in Detail
d-separation: revisited & details
Probability & Bayesian
Inference in & learning Bayesian networks
BN as AI tools and advantages
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 5

6. Bayesian Networks:
Introduction

7. Causality, Dependency
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
From correlation to causality
정성적방법
정량적방법
Granger causality index
http://www.google.com/trends/
Google의CausalImpact
R package for causal inference in time series
Official posting: http://google-
opensource.blogspot.kr/2014/09/causalimpact-new-open-
source-package.html
소개기사(영문): https://gigaom.com/2014/09/11/google-
has-open-sourced-a-tool-for-inferring-cause-from-
correlations/
7
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

8. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 8
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference
Assuming binary states for all the variables
Ex) Season: dry or rainy
Ex) Sprinkler: ON or OFF

9. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 9
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

10. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 10
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference
(directed acyclic graph)

11. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 11
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

12. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 12
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

13. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 13
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

14. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 14
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

15. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 15
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

16. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 16
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

17. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 17
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

18. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 18
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

19. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 19
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

20. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 20
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

21. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 21
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

22. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 22
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

23. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 23
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

24. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 24
Slide from K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal Inference

25. Correlation & Causation
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 25
Correlation does not imply causation
A chart that, correlates the number of
pirates with global temperature. The
two variables are correlated, but one
does not imply the other
a correlation between ice cream
consumption and crime, but
shows that the actual cause is
temperature.

26. Example
Designing a Bayesian Network
My own design of conditional probability tables
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 26
DRY 0.6
RAINY 0.4
SEASON
SEASON DRY RAINY
DRY 0.8 0.1
RAINY 0.2 0.9
SEASON DRY RAINY
YES 0.3 0.8
NO 0.7 0.2
WET YES NO
YES 0.8 0.1
NO 0.2 0.9
SPRINKLER RAIN
WET
SLIPPERY
SPRINKLER ON OFF
RAIN YES NO YES NO
YES 0.8 0.1 0.8 0.1
NO 0.2 0.9 0.2 0.9

27. Example
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 27
Designing a Bayesian Network
Tool: GeNIe(https://dslpitt.org/genie/)
GeNIe(Graphical Network Interface) is the graphical interface to SMILE,
a fully portable Bayesian inference engine in C++
Inference based on the designed Bayesian Network
Q1
A1

28. AI & Uncertainty

29. Probability
Probability plays a central role in modern pattern
recognition.
The main tool to deal uncertainties
All of the probabilistic inference and learning amount to
repeated application of the sum ruleand the product rule
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 29
Random Variables: variables + probability

30. Artificial Intelligence (AI)
The objective of AI is to build intelligent computers
We want intelligent, adaptive, robust behavior
Often hand programming not possible.
Solution? Get the computer to program itself, by
showing it examples of the behavior we want!
This is the learningapproach to AI.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 30
cat car

31. Artificial Intelligence (AI)
(Traditional) AI
Knowledge & reasoning; work with facts/assertions;
develop rules of logical inference
Planning: work with applicability/effects of actions;
develop searches for actions which achieve goals/avert
disasters.
Expert systems: develop by hand a set of rules for
examining inputs, updating internal states and
generating outputs
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 31

32. Artificial Intelligence (AI)
Probabilistic AI
emphasis on noisy measurements, approximation in hard
cases, learning, algorithmic issues.
The power of learning
Automatic system building
old expert systems needed hand coding of knowledge and
of output semantics
learning automatically constructs rules and supports all
types of queries
Probabilistic databases
traditional DB technology cannot answer queries about
items that were never loaded into the dataset
UAI models are like probabilistic databases
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 32

33. Uncertainty and Artificial Intelligence
(UAI)
Probabilistic methods can be used to:
make decisions given partial information about the world
account for noisy sensors or actuators
explain phenomena not part of our models
describe inherently stochastic behavior in the world
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 33

34. Other Names for UAI
Machine learning (ML), data mining, applied
statistics, adaptive (stochastic) signal processing,
probabilistic planning/reasoning...
Some differences:
Data mining almost always uses large data sets,
statistics almost always small ones
Data mining, planning, decision theory often have no
internal parameters to be learned
Statistics often has no algorithm to run!
ML/UAI algorithms are rarely online and rarely scale to
huge data (changing now).
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 34

35. Learning in AI
Learning is most useful
when the structure of the task is not well understood
but can be characterized by a dataset with strong
statistical regularity
Also useful in adaptive or dynamic situations when
the task (or its parameters) are constantly
changing
Currently, these are challenging topics of machine
learning and data mining research
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 35

36. Probabilistic AI
Let inputs=X, correct answers=Y, outputs of our
machine=Z
Learning: estimation of p(X, Y)
The central object of interest is the joint distribution
The main difficulty is compactly representing it and
robustly learning its shape given noisy samples
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 36

37. Probabilistic Graphical Models (PGMs)
Probabilistic graphical models represent large joint
distributions compactly using a set of “local”
relationshipsspecified by a graph
Each random variablein our model corresponds
to a graph node.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 37

38. Probabilistic Graphical Models (PGMs)
There are useful properties in using
probabilistic graphical models
A simple way to visualize the structure of a probabilistic
model
Insights into the properties of the model
Complex computations (for inference and learning) can
be expressed in terms of graphical manipulations
underlying mathematical expressions
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 38

39. 39
Directed graph vs. undirected graph
Both (probabilistic) graphical models
Specify a factorization(how to express the joint distribution)
Define a set of conditionalindependenceproperties
Parent -child
Local conditional distribution
Maximal clique
Potential function
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
Bayesian Networks (BN) Markov Random Field (MRF)

40. Bayesian Networks in Detail

41. © 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 41
(DAG)

42. Designing a Bayesian Network Model
TakeHeartII: Decision support system for clinical cardiovascular risk
assessment
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 42

43. Inference in a Bayesian Network Model
Given an assignment of a subset of variables (evidence) in a BN,
estimate the posterior distribution over another subset of unobserved
variables of interest.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 43
Inferences viewed as
message passing along the network

44. Bayesian Networks
The joint distribution defined by a graph is given by the
product of a conditional distribution of each node
conditioned on their parent nodes.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 44
????????????
1,??????
2,…,??????
7=
(??????�(??????
??????)denotes the set of parents of x
k)


K
k
kk
xPaxpp
1
))(|()(x
ex)
* Without given DAG structure, usual chain rule can be applied to get
the joint distribution. But computational cost is much higher.

45. Bayes’ Theorem
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 45(|)()
(|)
()
pXYpY
pYX
pX
 () (|)()
Y
pX pXYpY
Posterior
Likelihood
Prior
Normalizing
constant
posterior likelihood ×prior

46. Bayes’ Theorem
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 46
Figure from Figure 1. in (Adams, et all, 2013)
obtained from http://journal.frontiersin.org/article/10.3389/fpsyt.2013.00047/full

47. Bayesian Probabilities
-Frequentist vs. Bayesian
Likelihood:
Frequentist
w: a fixed parameter determined by ‘estimator’
Maximum likelihood: Error function =
Error bars: Obtained by the distribution of possible data sets
Bootstrap
Cross-validation
Bayesian
a probability distribution w: the uncertainty in the parameters
Prior knowledge
Noninformative(uniform) prior, Laplace correction in estimating priors
Monte Carlo methods, variationalBayes, EP
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 47(|)()
(|)
()
pp
p
p

ww
w
D
D
D (|)p wD log(|)p wD D
Thomas Bayes
(See an article ‘WHERE Do PROBABILITIES COME FROM?’ on page 491 in the textbook (Russell and Norvig, 2010) for more discussion)

48. Conditional Independence
Conditional independence simplifies both the structure of a model and
the computations
An important feature of graphical modelsis that conditional
independenceproperties of the joint distribution can be read directly
from the graphwithout having to perform any analytical manipulations
The general framework for this is called d-separation
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 48

49. Three example graphs –1
st
case
None of the variables are observed
The variable c is observed
The conditioned node ‘blocks’ the path from a to b,
causes a and b to become (conditionally) independent.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 49
Node c is tail-to-tail

50. Three example graphs –2
nd
case
None of the variables are observed
The variable c is observed
The conditioned node ‘blocks’ the path from a to b,
causes a and b to become (conditionally) independent.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 50
Node c is head-to-tail

51. Three example graphs –3
rd
case
None of the variables are observed
The variable c is observed
When node c is unobserved,
it ‘blocks’ the path and the variables a and b are independent.
Conditioning on c ‘unblocks’ the path and render a and b dependent.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 51
Node c is head-to-head

52. Three example graphs -Fuel gauge example
B –Battery, F-fuel, G-electric fuel gauge
Checking the fuel gauge
Checking the battery also has the meaning?
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 52
( Makes it more likely )
Makes it less likely than observation of fuel gauge only.
(rather unreliable fuel gauge)
(explaining away)

53. d-separation
Tail-to-tail node or head-to-tail node
Unless it is observed in which case it blocks a path, the
path is unblocked.
Head-to-head node
Blocks a path if is unobserved, but on the node, and/or
at least one of its descendants, is observed the path
becomes unblocked.
d-separation?
All paths are blocked.
The joint distribution will satisfy conditional
independence w.r.t. concerned variables.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 53

54. d-separation
(a) ais dependent to bgiven c
Head-to-head node eis unblocked, because a descendant c is in
the conditioning set.
Tail-to-tail node fis unblocked
(b) ais independent to bgiven f
Head-to-head node eis blocked
Tail-to-tail node fis blocked
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 54

55. d-separation
Another example of conditional independence and d-separation: i.i.d.
(independent identically distributed) data
Problem: finding posterior dist. for the mean of a univariate Gaussian dist.
Every path is blocked and so the observations D={x
1,…,x
N} are independent
given
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 55
(The observations are in general no longer independent!)
(independent)

56. d-separation
Naïve Bayes model
Key assumption: conditioned on the class z, the distribution of the
input variables x
1,…, x
Dare independent.
Input {x
1,…,x
N} with their class labels,
then we can fit the naïve Bayes model to the training data using
maximum likelihood assuming that the data are drawn
independently from the model.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 56

57. d-separation
Markov blanket or Markov boundary
When dealing with the conditional distribution of x
i , consider the
minimal set of nodes that isolates x
i from the rest of the graph.
The set of nodes comprising parents, children, co-parents is called
the Markov blanket.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 57
Co-parents
parents
children

58. Probability Distributions
Discrete variables
Beta, Bernoulli, binomial
Dirichlet, multinomial
…
Continuous variables
Normal (Gaussian)
Student-t
…
Exponential family & conjugacy
Many probability densities on xcan be represented as the same form
There are conjugate family of density functions having the same form of density
functions
Beta & binomial
Dirichlet& multinomial
Normal & Normal
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 58 (|)()()exp()
T
p hg x x ux
beta Dirichlet
binomial Gaussian
F
x
beta
binomial
Dirichlet
multinomial

59. Inference in Graphical Models
Inferencein graphical models
Given evidences (some nodes are clamped to observed values)
Wish to compute the posterior distributions of other nodes
Inference algorithms in graphical structures
Main idea: propagation of local messages
Exact inference
Sum-product algorithm, max-product algorithm, junction tree
algorithm
Approximate inference
Loopy belief propagation + message passing schedule
Variationalmethods, sampling methods (Monte Carlo methods)
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 59
A
BD
CE
ABD
BCD
CDE

60. Learning Parameters of
Bayesian Networks
Parameters
probabilities in conditional probability tables (CPTs) for all the
variables in the network
Learning parameters
Assuming that the structure is fixed, i.e. designed or learned.
We need data, i.e. observed instances
Estimation based on relative frequencies from data + belief
Example: coin toss. Estimation of ‘heads’ in various ways
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 60
SEASON DRY RAINY
YES ? ?
NO ? ?
RAIN
DRY ?
RAINY ?
SEASON
The principle of indifference: head and tail are equally probable
If we tossed a coin 10,000 times and it landed
heads 3373 times, we would estimate the
probability of heads to be about .3373
??????ℎ���??????=ൗ
1
2
1
2

61. Learning Parameters of
Bayesian Networks
Learning parameters (continued)
Estimation based on relative frequencies from data + belief
Example: A-match soccer game between Korea and Japan. How,
do you think, is it probable that Korean would win?
A: 0.85 (Korean), B: 0.3 (Japanese)
This probability is not a ratio, and it is not a relative
frequency because the game cannot be repeated many
times under the exact same conditions
Degree of beliefor subjective probability
Usual method
Estimate the probability distribution of a variable X
based on a relative frequencyand belief concerning a
relative frequency
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 61
3

62. Learning Parameters of
Bayesian Networks
Simple ‘counting’ solution (Bayesian point of view)
Parameter estimation of a single node
Assume local parameter independence
For a binary variable (for example, a coin toss)
prior: Beta distribution -Beta(a,b)
after we have observed mheads and N-mtails posterior -
Beta(a+m,b+N-m) and ????????????=ℎ���= ൗ
(??????+??????)
??????
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 62
(conjugacy of Beta and Binomial distributions)
beta
binomial

63. Learning Parameters of
Bayesian Networks
Simple ‘counting’ solution (Bayesian point of view)
For a multinomial variable (for example, a dice toss)
prior: Dirichletdistribution –Dirichlet(a
1,a
2, …, a
d)
????????????=??????=Τ
????????????
????????????=σ�
??????
Observing state i: Dirichlet(a
1,…,a
i
+1,…, a
d)
For an entire network
We simply iterate over its nodes
In the case of incomplete data
In real data, many of the variable values may be incorrect or missing
Usual approximating solution is given by Gibbs sampling or EM
(expectation maximization) technique
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 63
(conjugacy of Dirichletand Multinomial distributions)

64. Learning Parameters of
Bayesian Networks
Smoothing
Another viewpoint
Laplace smoothing or additive smoothing given observed counts for
dstates of a variable ??????=(??????
1,??????
2,…??????
??????)
From a Bayesianpoint of view, this corresponds to the expected
valueof the posterior distribution, using a symmetric Dirichlet
distributionwith parameter αas a prior.
Additive smoothing is commonly a component of naive Bayes
classifiers.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 64
????????????=??????=
??????
??????+??????
??????+??????�
??????=1,…,�,(??????=??????
1=??????
2=⋯??????
??????)

65. Learning the Graph Structure
Learning the graph structure itself from data requires
A space of possible structures
A measure that can be used to score each structure
From a Bayesian viewpoint
Tough points
Marginalization over latent variables => challenging computational
problem
Exploring the space of structures can also be problematic
The # of different graph structures grows exponentially with the # of
nodes
Usually we resort to heuristics
Local score based, global score based, conditional independence test based, …
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 65
: score for each model

66. Bayesian Networks as Tools for AI
Learning
Extracting and encoding knowledge from data
Knowledge is represented in Probabilistic relationship among variables
Causal relationship
Network of variables
Common framework for machine learning models
Supervised and unsupervised learning
Knowledge Representation & Reasoning
Bayesian networks can be constructed from prior knowledge alone
Constructed model can be used for reasoning based on probabilistic inference methods
Expert System
Uncertain expert knowledge can be encoded into a Bayesian network
DAG in a Bayesian network is hand-constructed by domain experts
Then the conditional probabilities were assessed by the expert, learned from data, or
obtained using a combination of both techniques.
Bayesian network-based expert systems are popular
Planning
In some different form, known as decision graphs or influence diagrams
We don’t cover about this direction
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 66

67. Advantages of Bayesian Networks for
Data Analysis
Ability to handle missing data
Because the model encodes dependencies among all variables
Learning causal relationships
Can be used to gain understanding about a problem domain
Can be used to predict the consequences of intervention
Having both causal and probabilistic semantics
It is an ideal representation for combining prior knowledge (which comes in
causal form) and data
Efficient and principled approach for avoiding the overfittingof data
By Bayesian statistical methods in conjunction with Bayesian networks
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 67
(summary from the abstract of D. Heckerman’s Tutorial on BN)
(Read ‘Introduction’ section for detailed explanations)

68. References
K. Mohan & J. Pearl, UAI ’12 Tutorial on Graphical Models for Causal
Inference
S. Roweis, MLSS ’05 Lecture on Probabilistic Graphical Models
Chapter 1, Chapter 2, Chapter 8 (Graphical Models), in Pattern
Recognition and Machine Learning by C.M. Bishop, 2006.
David Heckerman, A Tutorial on Learning with Bayesian Networks.
R.E. Neapolitan, Learning Bayesian Networks, Pearson Prentice Hall,
2004.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 68

69. More Textbooks and Courses
https://www.coursera.org/course/pgm:
Probabilistic Graphical Models by D. Koller
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 69

70. APPENDIX
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 70

71. Learning Parameters of
Bayesian Networks
Case 1: Xis a binary variable
F: beta distribution, X: Bernoulli or binomial distribution
Ex) F~ Beta(a,b), then ????????????=1=Τ
??????
??????(??????=�+�)
Case 2: Xis a multinomial variable
F: Dirichletdistribution, X: multinomial distribution
Ex) F~ Dirichlet(a
1,a
2, …,a
d), then ????????????=??????=Τ
??????
??????
????????????=σ�
??????
Laplace smoothing or additive smoothing given observed
frequencies ??????=(??????
1,??????
2,…??????
??????)
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 71
F
X
The probability distribution of Frepresents our belief
concerning the relative frequency with which Xequals k.
????????????=??????=
??????
??????+??????
??????+??????�
??????=1,…,�,(??????=??????
1=??????
2=⋯??????
??????)

72. Graphical interpretation of
Bayes’ theorem
Given structure:
We observe the value of y
Goal: infer the posterior distribution over x,
Marginal distribution : a prior over the latent
variable x
We can evaluate the marginal distribution
By Bayes’ theorem we can calculate
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 72(,)()(|)pxypxpyx ( | )p x y ()px ()py '
() (|')(')
x
py pyxpx (|)()
(|)
()
pyxpx
pxy
py

(a)
(b)
(c)

73. d-separation
Directed factorization
Filtering whether can be expressed in terms of the factorizationimplied by
the graph?
If we present to the filter the set of all possible distributions p(x) over the set
of variables X, then the subset of distributions that are passed by the filter
will be denoted DF(Directed Factorization)
Fully connected graph: The set DF will contain all possible distributions
Fully disconnected graph: The joint distributions which factorize into the
product of the marginal distributions over the variables only.
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 73

74. Gaussian distribution
© 2014-2015, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr 7422
21/2 2
11
(|,) exp ()
(2) 2
xx 
 

   

N 1
/ 2 1/ 2
1 1 1
( | , ) exp ( ) ( )
(2 ) | | 2
T
D


    

xμΣ xμΣxμ
Σ
N
Multivariate Gaussian