Bayes Classifiers. CAP5610 Machine Learning Instructor: Guo-Jun QI

Transcription

1 Bayes Classifiers CAP5610 Machine Learning Instructor: Guo-Jun QI

2 Recap: Joint distributions Joint distribution over Input vector X = (X 1, X 2 ) X 1 =B or B (drinking beer or not) X 2 = H or H (headache or not) Output vector Y = F or F (binary class) Input vector X Output Y F B H 0.4 F B H 0.1 F B H 0.17 F B H 0.2 F B H 0.05 F B H 0.05 F B H F B H 0.015

3 Prior distribution Prior for positive class P(Y=F) = = 0.13 F B H 0.4 F B H 0.1 F B H 0.17 F B H 0.2 F B H 0.05 Prior for negative class P(Y= F) = = 0.87 F B H 0.05 F B H F B H 0.015

4 Class-conditional distribution By Bayes Rule, P X 1, X 2 Y = F = P(X 1,X 2,Y=F) P(Y=F) P X 1 = B, X 2 = H Y = F = P(X 1 =B,X 2 =H,Y=F) P(Y=F) = = X 2 H H X 1 B ? B?? F B H 0.4 F B H 0.1 F B H 0.17 F B H 0.2 F B H 0.05 F B H 0.05 F B H F B H 0.015

5 Class-conditional distribution Bayes Rule P X 1, X 2 Y = F = P(X 1,X 2,Y= F) P(Y= F) P X 1 = B, X 2 = H Y = F = P(X 1 =B,X 2 =H,Y= F) P(Y= F) = = X 2 H H X 1 B ? B?? F B H 0.4 F B H 0.1 F B H 0.17 F B H 0.2 F B H 0.05 F B H 0.05 F B H F B H 0.015

6 Posterior distribution P Y = F X 1, X 2 = P X 1,X 2,Y=F P X 1,X 2 P Y = F X 1 = B, X 2 = H = P X 1 =B,X 2 =H,Y=F P X 1 =B,X 2 =H P(Y = F X 1, X 2 ) = P(X 1,X 2,Y= F) P(X 1,X 2 ) = F B H 0.4 F B H 0.1 F B H 0.17 F B H 0.2 F B H 0.05 F B H 0.05 F B H F B H 0.015

7 Prior, class-conditional, and posterior distribution Prior distribution for a class P(Y) has no input, the fraction of a particular class in a population What s the fraction of digit 9 among all digits [0-9]? What s the fraction of people who are infected with Flu?

8 Class-condition distribution Given a class, it is the distribution from which we can draw an example for this class. Digit 9 Flu X 2 H H X 1 B? B P(X 1,X 2 Y= F)

9 Posterior Distribution Posterior distribution for classes P(Y X 1,X 2 ) given an input X, what s the likelihood of a particular class? How likely is this image a digit 9? Digit 9?

10 Decision Theory Maximum A Posterior (MAP) Rule: given an input vector X, making an optimal decision about the class label (i.e., Y) in a certain sense Minimizing the classification error Case I: When P(Y=F X 1,X 2 )>P(Y= F X 1,X 2 ), X shall belong to F (i.e., X is infected with Flu) Case II: When P(Y=F X 1,X 2 )<P(Y= F X 1,X 2 ), X shall belong to F (i.e., X is not infected with Flu) Proof: MAP rule gives the minimal classification error.

11 Proof Decision region defines a region in the feature space such that every point in this region belongs to a particular class. X 2 F F R 1 R 2 F Decision boundary X 1

12 Proof p error = R1 p X, Y = C 2 dx + R2 p(x, Y = C 1 ) dx = R 1 p Y = C 2 X p(x) dx + R 2 p Y = C 1 X p(x) dx For each X, it either belongs to R1 or R2; to minimize the error rate, it shall be assigned to the region with a smaller posterior probability.

13 Likelihood Ratio Maximum A Posterior rule Case I: When P(Y=F X 1,X 2 )>P(Y= F X 1,X 2 ), X shall belong to F (i.e., X is infected with Flu) Case II: When P(Y=F X 1,X 2 )<P(Y= F X 1,X 2 ), X shall belong to F (i.e., X is infected with Flu) Likelihood Ratios f X = P(Y= F X 1,X 2 ) P(Y= F X 1,X 2 ) Where f(x) > 1, X belongs to F, otherwise X belongs to F

14 Discriminate Function Given an input X, a discriminative function decides its class by comparing f(x) with a certain threshold. f X > 1, X F < 1, X F Discriminative function does not need to be positive, or its threshold does not need to be 1 either.

15 A linear discriminate function f(x) = X 1 + X 2-1with threshold 0. (0,1) X 2 f(x) > 0 R 1 Linear decision boundary R 2 f(x) < 0 (1,0) X 1

16 Bayes Error Bayes error is the minimal error that is made by MAP rule. It is the lowest bound of error rate that can be achieved by any classifier p error X = p Y = C 1 X, if P Y = C 2 X > P(Y = C 1 X) p Y = C 2 X, if P Y = C 1 X > P(Y = C 2 X) = min{p Y = C 1 X, p Y = C 2 X }

17 Nearest Neighbor Error The error made by nearest neighbor classifier (1-NN) is smaller than twice Bayes error. Given a example X, its nearest neighbor is X NN ; the true class of X is Y, and test the true class of X NN is Y NN. p NN error X, X NN = p Y = C 1, Y NN = C 2 X, X NN + p Y = C 2, Y NN = C 1 X, X NN = p Y = C 1 X p(y NN = C 2 X NN + p Y = C 2 X p(y NN = C 1 X NN When the size of training set is large enough (approaching to infinity), X NN will also approach to X p NN error X = 2p Y = C 1 X p(y = C 2 X

18 Bayes error and NN asymptotic error Bayes error: p error X = min{p Y = C 1 X, p Y = C 2 X } NN asymptotic error: p NN error X = 2p Y = C 1 X p(y = C 2 X p NN error X < 2p error X

19 Bayesian Classifier Comparing the posterior distribution Given an input feature vector X, Y = C 1, if P Y = C 2 X > P(Y = C 1 X) C 2, if P Y = C 1 X > P(Y = C 2 X) where P Y = C i X P X Y = Ci P Y = Ci, i = 1,2

20 Practical Issue Prior distribution P Y = C i, i = 1,2 Counting the fraction of two classes in training set Class-conditional distribution P X Y = C i, i = 1,2 Modeled from the training examples belonging to two classes Four training examples X 1 (Drinking beer) X 2 (Headache) Y (Flu) P X = (0,1) Y = 1 = P X = (1,1) Y = 1 = P X = (1,0) Y = 1 = P X = (0,0) Y = 1 = #(X = 0,1, Y = 1) #(Y = 1) #(X = 1,1, Y = 1) #(Y = 1) #(X = 1,0, Y = 1) #(Y = 1) #(X = 0,0, Y = 1) #(Y = 1) = 1 2 = 1 2 = 0 2 = 0 2

21 N attributes of feature vector Input vector: X=(X 1,X 2,,X N ) Estimate P(X Y=C i ), how many examples suffice to do estimation? Assume X is binary vector, then at least 2 N examples are required to ensure that each possible assignment of binary attributes to X has one training example. N=20, 2 N = 1,048,576 N=30, 2 N = 1,073,741,824 In MNIST, N = 28X28 (pixel), 2 N = 1.01X10 236, let alone X is continuous vector

22 Naive Bayes Assume attributes are independent of each other given a class C i, i=1,2 P X 1, X 2,, X N Y = C i = P X 1 Y = Ci P X 2 Y = C i P(XN Y = C i ) Each P X n Y = C i can be estimated independently Only 2 training examples are required at least to estimate each P X n Y = C i binary At least 2N training examples can estimate the joint distribution. if X n is

23 Naive Bayes P X 1 = 0 Y = 1 = #(X 1 = 0, Y = 1) #(Y = 1) P X 1 = 1 Y = 1 =? = 1 2 X 1 (Drinking beer) X 2 (Headache) Y (Flu) An exercise: complete the Naive Bayes (not homework)

24 Summary Recap prior distribution, class-conditional distribution, posterior distribution Maximum A Posterior (MAP) Rule to decide the class assigned to each input vector X Likelihood Ratio, and discriminant function Decision Boundary and Region Practical Issue: Estimate prior distribution and class-conditional distribution from training example Naive Bayes