Bayesian Networks

Motivation The conditional independence assumption made by naïve Bayes classifiers may seem to rigid, especially for classification problems in which the attributes are somewhat correlated. We talk today for a more flexible approach for modeling the conditional probabilities.

Naïve Bayes and Correlated Attrs Two binary attributes A and B, and a binary class Y. A distribution –P(A=0 | Y=0) = 0.4P(A=1 | Y=0) = 0.6 –P(A=0 | Y=1) = 0.6P(A=1 | Y=1) = 0.4 B is perfectly correlated with A when Y=0, but not when Y=1 –P(B=0 | Y=0) = 0.4P(B=1 | Y=0) = 0.6 –P(B=0 | Y=1) = 0.5P(B=1 | Y=1) = 0.5 Also, let’s assume –P(Y=0) = P(Y=1) = 0.5

Naïve Bayes and Correlated Attrs Now, we are given a new record with A=0 and B=0. Using Naïve Bayes we have: P(Y=0 | A=0, B=0) =  * P(A=0 | Y=0) * P(B=0 | Y=0) * P(Y=0) =  *.4 *.4 *.5 =  * 0.08 P(Y=1 | A=0, B=0) =  * P(A=0 | Y=1) * P(B=0 | Y=1) * P(Y=1) =  *.6 *.5 *.5 =  *.15 So, we predict Y=1.

Naïve Bayes and Correlated Attrs However, since A and B are perfectly correlated (when Y=0) we have that: P(A=0, B=0 | Y=0) = P(A=0 | Y=0) = 0.4 Thus, P(Y=0 | A=0, B=0) =  * P(A=0, B=0 | Y=0) * P(Y=0) =  * 0.4 * 0.5 =  * 0.2 which is greater than P(Y=1 | A=0, B=0) =  * P(A=0 | Y=1) * P(B=0 | Y=1) * P(Y=1) =  *.6 *.5 *.5 =  *.15 So, the record should have been classified as class 0.

Bayesian networks A simple, graphical notation for conditional independence assertions. Syntax: a set of nodes, one per variable (attribute) a directed, acyclic graph (link means: "directly influences") a conditional distribution for each node given its parents: P (X i | Parents (X i )) The conditional distribution is represented as a conditional probability table (CPT) giving the distribution over X i for each combination of parent values.

Example (Perls’ example) I'm at work, neighbor John calls to say my alarm is ringing, but neighbor Mary doesn't call. Sometimes it's set off by minor earthquakes. Is there a burglar? John always calls when he hears the alarm, but sometimes confuses the telephone ringing with the alarm. Mary likes rather loud music and sometimes misses the alarm. Variables: Burglary, Earthquake, Alarm, JohnCalls, MaryCalls Network topology reflects "causal" knowledge: –A burglar can set the alarm off –An earthquake can set the alarm off –The alarm can cause Mary to call –The alarm can cause John to call

Example cont’d The topology shows that burglary and earthquakes directly affect the probability of alarm, but whether Mary or John call depends only on the alarm. Thus our assumptions are that they don’t perceive any burglaries directly, and they don’t confer before calling. To save space, some of the probabilities have been omitted from the diagram. The omitted probabilities can be recovered by noting that P(X = x) = 1 - P(X =  x) and P(X =  x|Y) = 1 - P(X=  x|Y), where  x denotes the opposite outcome of x.

Semantics e.g., P(j  m  a   b   e) = P(j | a) P(m | a) P(a |  b,  e) P(  b) P(  e) = … Suppose we have the variables X 1,…,X n. The probability for them to have the values x 1,…,x n respectively is P(x n,…,x 1 ): P(x n,…,x 1 ): is short for P(X n =x n,…, X n = x 1 ):

Inference in Bayesian Networks The basic task for a probabilistic inference system is to compute the posterior probability for a query variable (class attribute), given some observed event –that is, some assignment of values to a set of evidence variables (some of the other attributes). Notation: –X denotes query variable –E denotes the set of evidence variables E 1,…,E m, and e is a particular event, i.e. an assignment to the variables in E. –Y will denote the set of the remaining variables (hidden variables). A typical query asks for the posterior probability P(x|e 1,…,e m ) E.g. We could ask: What’s the probability of a burglary if both Mary and John call, P(burglary | johhcalls, marycalls)?

Classification Suppose, we are given for the evidence variables E 1,…,E m, their values e 1,…,e m, and we want to predict whether the query variable X has the value x or not. For this we compute and compare the following: However, how do we compute: What about the hidden variables Y 1,…,Y k ?

Inference by enumeration Example: P(burglary | johhcalls, marycalls)? (Abbrev. P(b|j,m)) In general:

Numerically… P(b | j,m) =  P(b)  a P(j|a)P(m|a)  e P(a|b,e)P(e) = …=  * P(  b | j,m) =  P(  b)  a P(j|a)P(m|a)  e P(a|  b,e)P(e) = …=  * P(B | j,m) =  =.

P(b | j,m) P(b | j,m) =  P(b)  a P(j|a)P(m|a)  e P(a|b,e)P(e) =  P(b)  a P(j|a)P(m|a)(P(a|b,e)P(e) + P(a|b,  e)P(  e)) =  P(b)( P(j|a)P(m|a)( P(a|b,e)P(e) + P(a|b,  e)P(  e) ) + P(j|  a)P(m|  a)( P(  a|b,e)P(e) + P(  a|b,  e)P(  e) )) =  *.001*(.9*.7*(.95* *.998) +.05*.01*(.05* *.998) ) =  *.00059

P(  b | j,m) P(  b | j,m) =  P(  b)  a P(j|a)P(m|a)  e P(a|  b,e)P(e) =  P(  b)  a P(j|a)P(m|a)(P(a|  b,e)P(e) + P(a|  b,  e)P(  e)) =  P(  b)( P(j|a)P(m|a)( P(a|  b,e)P(e) + P(a|  b,  e)P(  e) ) + P(j|  a)P(m|  a)( P(  a|  b,e)P(e) + P(  a|  b,  e)P(  e) )) =  *.999*(.9*.7*(.29* *.998) +.05*.01*(.71* *.998) ) =  *.0015  = 1/( ) = P(b | j,m) = * =.28 P(  b | j,m) = *.0015 =.72

Constructing Bayesian networks 1. Choose an ordering of variables X 1, …,X n 2. For i = 1 to n –add X i to the network –select parents from X 1, …,X i-1 such that P(X i | Parents(X i )) = P(X i | X 1,... X i-1 ) This choice of parents guarantees: P(X 1, …,X n ) =  i =1 P(X i | X 1, …, X i-1 ) (chain rule) =  i =1 P(X i | Parents(X i )) (by construction) Choosing the parents from X 1, …, X i-1 is done by domain human experts.

The ordering of variables is very important. E.g. suppose we choose the ordering M, J, A, B, E Adding MaryCalls: No parents P (J|M) = P (J)? Is P(John calling) independent of P(Mary calling)? Clearly not, since, on any given day, if Mary called, then the probability that John called is much better than the background probability that he called. So, we add a link from MaryCalls to JohnCalls. Example

Suppose we choose the ordering M, J, A, B, E Adding the A (Alarm) node: Is P(A | J, M) = P(A | J)? P(A | J, M) = P(A)? No. Clearly, if both call, it’s more likely that the alarm has gone off that if just one or neither call, so we need both MaryCalls and JohnCalls as parents. Example

Suppose we choose the ordering M, J, A, B, E Adding B (Burglary) node: Is P(B | A, J, M) = P(B | A)? P(B | A, J, M) = P(B)? Yes for the first. No for the second. If we know the alarm state, then the call from John or Mary might give us information about the phone ringing or Mary’s music, but not about burglary. So, we need just Alarm as parent. Example

Suppose we choose the ordering M, J, A, B, E Adding E (Earthquake) node: Is P(E | B, A,J, M) = P(E | A)? P(E | B, A, J, M) = P(E | A, B)? No for the first. Yes for the second. If the alarm is on, it is more likely that there has been an earthquake. But if we know there has been a burglary, then that explains the alarm, and the probability of an earthquake would be only slightly above normal. Hence we need both Alarm and Burglary as parents. Example

Example cont’d So, the n etwork is less compact if we go non-causal: = 13 numbers needed instead of 10 if we go in causal direction. Deciding conditional independence is harder in noncausal directions Causal models and conditional independence seem hardwired for humans!