1 Machine Learning: Lecture 6 Bayesian Learning (Based on Chapter 6 of Mitchell T.., Machine Learning, 1997)

2 An Introduction §Bayesian Decision Theory came long before Version Spaces, Decision Tree Learning and Neural Networks. It was studied in the field of Statistical Theory and more specifically, in the field of Pattern Recognition. §Bayesian Decision Theory is at the basis of important learning schemes such as the Naïve Bayes Classifier, Learning Bayesian Belief Networks and the EM Algorithm. §Bayesian Decision Theory is also useful as it provides a framework within which many non-Bayesian classifiers can be studied (See [Mitchell, Sections 6.3, 4,5,6]).

3 Bayes Theorem §Goal: To determine the most probable hypothesis, given the data D plus any initial knowledge about the prior probabilities of the various hypotheses in H. §Prior probability of h, P(h): it reflects any background knowledge we have about the chance that h is a correct hypothesis (before having observed the data). §Prior probability of D, P(D): it reflects the probability that training data D will be observed given no knowledge about which hypothesis h holds. §Conditional Probability of observation D, P(D|h): it denotes the probability of observing data D given some world in which hypothesis h holds.

4 Bayes Theorem (Cont’d) §Posterior probability of h, P(h|D): it represents the probability that h holds given the observed training data D. It reflects our confidence that h holds after we have seen the training data D and it is the quantity that Machine Learning researchers are interested in. §Bayes Theorem allows us to compute P(h|D): P(h|D)=P(D|h)P(h)/P(D)

5 Maximum A Posteriori (MAP) Hypothesis and Maximum Likelihood §Goal: To find the most probable hypothesis h from a set of candidate hypotheses H given the observed data D. §MAP Hypothesis, h MAP = argmax h  H P(h|D) = argmax h  H P(D|h)P(h)/P(D) = argmax h  H P(D|h)P(h) §If every hypothesis in H is equally probable a priori, we only need to consider the likelihood of the data D given h, P(D|h). Then, h MAP becomes the Maximum Likelihood, h ML = argmax h  H P(D|h)P(h)

6 An Example §Two alternative hypotheses l The patient has a particular form of cancer l The patient does not §Observed Data l Positive/negative of a particular lab test §Prior knowledge l Population of people have this disease

7 Bayes Theorem and Concept Learning §Assumption l Training data D is noise free. l The target concept c is contained in H l No a priori reason to believe that any hypothesis is more probable than any other

8 Bayes Theorem and Concept Learning See Fig. 6.1 in text If P(h)=1/|H| and if P(D|h)=1 if D is consistent with h, and 0 otherwise, then every hypothesis in the version space resulting from D is a MAP hypothesis.

9 MAP Hypotheses and Consistent Learners §Every hypothesis consistent with D is a MAP hypothesis. §Consistent Learner l It outputs a hypothesis that commits zero errors over training examples. l Every consistent learner outputs a MAP hypothesis if we assume a uniform prior probability distribution over H and if we assume deterministic, noise free data. l Example: Find-S algorithm outputs MAP hypotheses that are maximally specific members of version space. Are there other probability distributions for P(h) and P(D|h) ? –Yes.

10 MAP Hypotheses and Consistent Learners §Bayesian framework allows one way to characterize the behavior of learning algorithms (e.g. Find-S), even when the learning algorithm does not explicitly manipulate probabilities.

11 ML and Least-Squared Error Hypotheses §Under certain assumptions regarding noise in the data, minimizing the mean squared error (what common neural nets do) corresponds to computing the maximum likelihood hypothesis.

12 ML and Least-Squared Error Hypotheses

13 ML and Least-Squared Error Hypotheses §Assumptions l MAP becomes ML when P(h) is equally probable. l Normal Distribution of error l Noise in target values but noise free for attribute values Target value: Weight Attributes: Age, Height

14 ML Hypotheses for Predicting Probability §Binary classifier of neural net l Training data l Single output node

15 ML Hypotheses for Predicting Probability

16 ML Hypotheses for Predicting Probability §Gradient Search in single layer NN §Summary: l Minimize sum of squared error seeks the ML hypothesis under assumption that training data can be modeled by Normal distributed noise added to the target function value. l The rule that minimizes cross entropy seeks the ML hypothesis under the assumption that the observed Boolean value is the probabilistic function of input instance.

17

18 Bayes Optimal Classifier §One great advantage of Bayesian Decision Theory is that it gives us a lower bound on the classification error that can be obtained for a given problem. §Bayes Optimal Classification: The most probable classification of a new instance is obtained by combining the predictions of all hypotheses, weighted by their posterior probabilities: argmax vj  V  hi  H P(v h |h i )P(h i |D) where V is the set of all the values a classification can take and v j is one possible such classification. §Unfortunately, Bayes Optimal Classifier is usually too costly to apply! ==> Naïve Bayes Classifier

19

20 Gibbs algorithm §Algorithm l 1. Choose h from H, according to the posterior probability distribution over H l 2. Use h to predict the classification of x §Validity of Gibbs algorithm l Haussler, 1994 l Error(Gibbs algorithm)< 2*Error(Bayes optimal classifier)

21 Naïve Bayes Classifier §Let each instance x of a training set D be described by a conjunction of n attribute values and let f(x), the target function, be such that f(x)  V, a finite set. §Bayesian Approach: v MAP = argmax vj  V P(v j |a 1,a 2,..,a n ) = argmax vj  V [P(a 1,a 2,..,a n |v j ) P(v j )/P(a 1,a 2,..,a n )] = argmax vj  V [P(a 1,a 2,..,a n |v j ) P(v j ) §Naïve Bayesian Approach: We assume that the attribute values are conditionally independent so that P(a 1,a 2,..,a n |v j ) =  i P(a i |v j ) [and not too large a data set is required.] § Naïve Bayes Classifier: v NB = argmax vj  V P(v j )  i P(a i |v j )

22 An Illustrative Example §(outlook=sunny,temperature=cool,humidity=high,wind=str ong) §P(wind=strong|playTennis=yes)=3/9=.33 §P(wind=strong|PlayTennis=no)=3/5=.60 §P(yes)P(sunny|yes)P(cool|yes)P(high|yes)P(strong|yes)= §P(no)P(sunny|no)P(cool|no)P(high|no)P(strong|no)=.0206 §v NB = no

23 Naïve Bayes Classifier Estimating Probabilities P(a i |v j ) m-estimate of probability = –m : equivalent sample size, p : prior estimate of probability –n c : number of instances a i among v j –n: number of instances with value v j Example: Table 3.2

24 An Example: Learning to classify text §Independent Assumption l Incorrect assumption but no other choice l Example: machine learning l Impractical: 2 target values*111 word positions*50000 words l Position Independent Assumption 2 target values*50000 words See Algorithm at page 183 l Learn_Bayes_Text/Classify_Bayes_Text

25 Bayesian Belief Networks §The Bayes Optimal Classifier is often too costly to apply. §The Naïve Bayes Classifier uses the conditional independence assumption to defray these costs. However, in many cases, such an assumption is overly restrictive. §Bayesian belief networks provide an intermediate approach which allows stating conditional independence assumptions that apply to subsets of the variable.

26 Conditional Independence § We say that X is conditionally independent of Y given Z if the probability distribution governing X is independent of the value of Y given a value for Z. §i.e., (  x i,y j,z k ) P(X=x i |Y=y j,Z=z k )=P(X=x i |Z=z k ) §or, P(X|Y,Z)=P(X|Z) §This definition can be extended to sets of variables as well: we say that the set of variables X 1 …X l is conditionally independent of the set of variables Y 1 …Y m given the set of variables Z 1 …Z n, if P(X 1 …X l |Y 1 …Y m,Z 1 …Z n )=P(X 1 …X l |Z 1 …Z n )

27 Representation in Bayesian Belief Networks Storm BusTourGroup Lightning Campfire ThunderForestFire Each node is asserted to be conditionally independent of its non-descendants, given its immediate parents Associated with each node is a conditional probability table, which specifies the conditional distribution for the variable given its immediate parents in the graph

28 Inference in Bayesian Belief Networks §A Bayesian Network can be used to compute the probability distribution for any subset of network variables given the values or distributions for any subset of the remaining variables. §Unfortunately, exact inference of probabilities in general for an arbitrary Bayesian Network is known to be NP-hard. §In theory, approximate techniques (such as Monte Carlo Methods) can also be NP-hard, though in practice, many such methods were shown to be useful.

29 Learning Bayesian Belief Networks 3 Cases: 1. The network structure is given in advance and all the variables are fully observable in the training examples. ==> Trivial Case: just estimate the conditional probabilities. 2. The network structure is given in advance but only some of the variables are observable in the training data. ==> Similar to learning the weights for the hidden units of a Neural Net: Gradient Ascent Procedure 3. The network structure is not known in advance. ==> Use a heuristic search or constraint-based technique to search through potential structures.

30 The EM Algorithm: Learning with unobservable relevant variables. §Example:Assume that data points have been uniformly generated from k distinct Gaussian with the same known variance. The problem is to output a hypothesis h= that describes the means of each of the k distributions. In particular, we are looking for a maximum likelihood hypothesis for these means. §We extend the problem description as follows: for each point x i, there are k hidden variables z i1,..,z ik such that z il =1 if x i was generated by normal distribution l and z iq = 0 for all q  l.

31 The EM Algorithm (Cont’d) §An arbitrary initial hypothesis h= is chosen. §The EM Algorithm iterates over two steps: l Step 1 (Estimation, E): Calculate the expected value E[zij] of each hidden variable zij, assuming that the current hypothesis h= holds. l Step 2 (Maximization, M): Calculate a new maximum likelihood hypothesis h’=, assuming the value taken on by each hidden variable z ij is its expected value E[z ij ] calculated in step 1. Then replace the hypothesis h= by the new hypothesis h’= and iterate. The EM Algorithm can be applied to more general problems