1. Marginalization & Conditioning Marginalization (summing out): for any sets of variables Y and Z: Conditioning(variant of marginalization):

2. Example of Marginalization Using the full joint distribution P(cavity)= P(cavity, toothache, catch) + P(cavity, toothache,  catch) + P(cavity,  toothache, catch) + P(cavity,  toothache,  catch) = 0.108 +0.012 + 0.072 + 0.008 = 0.2

3. Inference By Enumeration using Full Joint Distribution Let X be a random variable about which we want to know its probabilities, given some evidence (values e for a set E of other variables). Let the remaining (unobserved, so-called hidden) variables be Y. The query is P(X|e), and it can be answered using the full joint distribution by

4. Example of Inference By Enumeration using Full Joint Distribution

5. Independence Propositions a and b are independent if and only if Equivalently (by product rule): Equivalently:

6. Illustration of Independence We know (product rule) that

7. Illustration continued Allows us to represent a 32-element table for full joint on Weather, Toothache, Catch, Cavity by an 8-element table for the joint of Toothache, Catch, Cavity, and a 4-element table for Weather. If we add a Boolean variable X to the 8- element table, we get 16 elements. A new 2-element table suffices with independence.

8. Difficulty with Bayes’ Rule with More than Two Variables

9. Conditional Independence X and Y are conditionally independent given Z if and only if P(X,Y|Z) = P(X|Z) P(Y|Z). Y 1,…,Y n are conditionally independent given X 1,…,X m if and only if P(Y 1,…,Y n |X 1,…,X m )= P(Y 1 |X 1,…,X m ) P(Y 2 |X 1,…,X m ) … P(Y m |X 1,…,X m ). We’ve reduced 2 n 2 m to 2n2 m. Additional conditional independencies may reduce 2 m.

10. Conditional Independence As with absolute independence, the equivalent forms of X and Y being conditionally independent given Z can also be used: P(X|Y, Z) = P(X|Z) and P(Y|X, Z) = P(Y|Z)

11. Benefits of Conditional Independence Allows probabilistic systems to scale up (tabular representations of full joint distributions quickly become too large.) Conditional independence is much more commonly available than is absolute independence.

12. Decomposing a Full Joint by Conditional Independence Might assume Toothache and Catch are conditionally independent given Cavity: P(Toothache,Catch|Cavity) = P(Toothache|Cavity) P(Catch|Cavity). Then P(Toothache,Catch,Cavity) = [product rule] P(Toothache,Catch|Cavity) P(Cavity) = [conditional independence] P(Toothache|Cavity) P(Catch|Cavity) P(Cavity).

13. Naive Bayes Algorithm Let Fi be the i-th feature having valuej and Out be the target feature. We can use training data to estimate: P(F i = v j ) P(F i = v j | Out = True) P(F i = v j | Out = False) P(Out = True) P(Out = False)

14. Naive Bayes Algorithm For a test example described by F 1 = v 1,..., F n = v n, we need to compute: P(Out = True | F 1 = v 1,..., F n = v n ) Applying Bayes rule: P(Out = True | F 1 = v 1,..., F n = v n ) = P(F 1 = v 1,..., F n = v n | Out = True) P(Out = True) _______________________________________ P(F 1 = v 1,..., F n = v n )

15. Naive Bayes Algorithm By independence assumption: P(F 1 = v 1,..., F n = v n ) = P(F 1 = v 1 )x...x P(F n = v n ) This leads to conditional independence: P(F 1 = v 1,..., F n = v n | Out = True) = P(F 1 = v 1 | Out = True) x...x P(F n = v n | Out = True)

16. Naive Bayes Algorithm P(Out = True | F 1 = v 1,..., F n = v n ) = P(F 1 = v 1 | Out = True) x...x P(F n = v n | Out = True)x P(Out = True) _______________________________________ P(F 1 = v 1 )x...x P(F n = v n ) All terms are computed using the training data! Works well despite of strong assumptions(see [Domingos and Pazzani MLJ 97]) and thus provides a simple benchmark testset accuracy for a new data set

17. Bayesian Networks: Motivation Although the full joint distribution can answer any question about the domain it can become intractably large as the number of variable grows. Specifying probabilities for atomic events is rather unnatural and may be very difficult. Use a graphical representation for which we can more easily investigate the complexity of inference and can search for efficient inference algorithms.

18. Bayesian Networks Capture independence and conditional independence where they exist, thus reducing the number of probabilities that need to be specified. It represents dependencies among variables and encodes a concise specification of the full joint distribution.

19. A Bayesian Network is a... Directed Acyclic Graph (DAG) in which … … the nodes denote random variables … each node X has a conditional probability distribution P(X|Parents(X)). The intuitive meaning of an arc from X to Y is that X directly influences Y.

20. Additional Terminology If X and its parents are discrete, we can represent the distribution P(X|Parents(X)) by a conditional probability table (CPT) specifying the probability of each value of X given each possible combination of settings for the variables in Parents(X). A conditioning case is a row in this CPT (a setting of values for the parent nodes). Each row must sum to 1.

21. Bayesian Network Semantics A Bayesian Network completely specifies a full joint distribution over its random variables, as below -- this is its meaning. P In the above, P(x 1,…,x n ) is shorthand notation for P(X 1 =x 1,…,X n =x n ).

22. Inference Example What is probability alarm sounds, but neither a burglary nor an earthquake has occurred, and both John and Mary call? Using j for John Calls, a for Alarm, etc.:

23. Chain Rule Generalization of the product rule, easily proven by repeated application of the product rule. Chain Rule:

24. Chain Rule and BN Semantics

25. Example of the Key Property The following conditional independence holds: P (MaryCalls |JohnCalls, Alarm, Earthquake, Burglary) = P(MaryCalls | Alarm)

26. Procedure for BN Construction Choose relevant random variables. While there are variables left:

27. Principles to Guide Choices Goal: build a locally structured (sparse) network -- each component interacts with a bounded number of other components. Add root causes first, then the variables that they influence.