Naïve Bayes Introduction to Machine Learning. Matt Gormley Lecture 3 September 14, Readings: Mitchell Ch Murphy Ch.

Transcription

1 School of Computer Science Introduction to Machine Learning aïve Bayes Readings: Mitchell Ch Murphy Ch. 3 Matt Gormley Lecture 3 September 14,

2 Homewor 1: due 9/26/16 Project Proposal: due 10/3/16 start early! Reminders 2

3 Outline Bacground: Maximum lielihood estimation (MLE) Maximum a posteriori (MAP) estimation Example: Exponential distribution Generative Models Model 0: ot- so- naïve Model aïve Bayes aïve Bayes Assumption Model 1: Bernoulli aïve Bayes Model 2: Multinomial aïve Bayes Model 3: Gaussian aïve Bayes Model 4: Multiclass aïve Bayes Smoothing Add- 1 Smoothing Add- λ Smoothing MAP Estimation (Beta Prior) 3

4 MLE vs. MAP Suppose we have data D = {x (i) } i=1 MLE = p( (i) ) Maximum Lielihood Estimate (MLE) i=1 4

5 Bacground: MLE Example: MLE of Exponential Distribution pdf of Exponential( ): f(x) = e x Suppose X i Exponential( ) for 1 i. Find MLE for data D = {x (i) } i=1 First write down log-lielihood of sample. Compute first derivative, set to zero, solve for. Compute second derivative and chec that it is concave down at MLE. 5

6 Bacground: MLE Example: MLE of Exponential Distribution First write down log-lielihood of sample. ( )= = = i=1 i=1 i=1 = ( ) f(x (i) ) (1) ( ( x (i) )) (2) ( )+ x (i) (3) i=1 x (i) (4) 6

7 Bacground: MLE Example: MLE of Exponential Distribution Compute first derivative, set to zero, solve for. d ( ) d = d d ( ) = i=1 MLE = i=1 x (i) (1) x (i) =0 (2) i=1 x(i) (3) 7

8 Bacground: MLE Example: MLE of Exponential Distribution pdf of Exponential( ): f(x) = e x Suppose X i Exponential( ) for 1 i. Find MLE for data D = {x (i) } i=1 First write down log-lielihood of sample. Compute first derivative, set to zero, solve for. Compute second derivative and chec that it is concave down at MLE. 8

9 MLE vs. MAP Suppose we have data D = {x (i) } i=1 MLE = MAP = i=1 i=1 p( (i) ) p( (i) )p( ) Maximum Lielihood Estimate (MLE) Maximum a posteriori (MAP) estimate Prior 9

10 Generative Models Specify a generative story for how the data was created (e.g. roll a weighted die) Given parameters (e.g. weights for each side) for the model, we can generate new data (e.g. roll the die again) Typical learning approach is MLE (e.g. find the most liely weights given the data) 10

11 Features Suppose we want to represent a document (M words) as a vector (vocabulary size V ) How should we do it? Option 1: Integer vector (word IDs) =[x 1,x 2,...,x M ] where x m {1,...,K} a word id. Option 2: Binary vector (word indicators) =[x 1,x 2,...,x K ] where x {0, 1} is a boolean. Option 3: Integer vector (word counts) =[x 1,x 2,...,x K ] where x Z + is a positive integer. 11

12 Today s Goal To define a generative model of s of two different classes (e.g. spam vs. not spam) 12

13 Spam ews The Economist The Onion 13

14 Model 0: ot- so- naïve Model? Generative Story: 1. Flip a weighted coin (Y) 2. If heads, roll the red many sided die to sample a document vector (X) from the Spam distribution 3. If tails, roll the blue many sided die to sample a document vector (X) from the ot- Spam distribution P (X 1,...,X K,Y)=P (X 1,...,X K Y )P (Y ) This model is computationally naïve! 14

15 Model 0: ot- so- naïve Model? Generative Story: 1. Flip a weighted coin (Y) 2. If heads, sample a document ID (X) from the Spam distribution 3. If tails, sample a document ID (X) from the ot- Spam distribution P (X, Y )=P (X Y )P (Y ) This model is computationally naïve! 15

16 Model 0: ot- so- naïve Model? Flip weighted coin If HEADS, roll red die y x 1 x 2 x 3 x K If TAILS, roll blue die Each side of the die is labeled with a document vector (e.g. [1,0,1,,1])

17 Outline Bacground: Maximum lielihood estimation (MLE) Maximum a posteriori (MAP) estimation Example: Exponential distribution Generative Models Model 0: ot- so- naïve Model aïve Bayes aïve Bayes Assumption Model 1: Bernoulli aïve Bayes Model 2: Multinomial aïve Bayes Model 3: Gaussian aïve Bayes Model 4: Multiclass aïve Bayes Smoothing Add- 1 Smoothing Add- λ Smoothing MAP Estimation (Beta Prior) 17

18 aïve Bayes Assumption Conditional independence of features: P (X 1,...,X K,Y)=P(X 1,...,X K Y )P (Y ) = K =1 P (X Y ) P (Y ) 18

19 C P(C) A C P(A C) B C P(B C) Estimating a joint from conditional probabilities P(A, B C) = P(A C)* P(B C) a, bc : P(A = a B = b C = c) = P(A = a C = c)* P(B = b C = c) A B C P(A,B,C)

20 C P(C) A C P(A C) B 0.5 C P(B C) D C P(D C) Estimating a joint from conditional probabilities A B D C P(A,B,D,C)

21 Assuming conditional independence, the conditional probabilities encode the same information as the joint table. They are very convenient for estimating P( X 1,,X n Y)=P( X 1 Y)* *P( X n Y) They are almost as good for computing P(Y X 1,..., X n ) = P(X..., X 1, n Y )P(Y ) P(X 1,..., X n ) x,y : P(Y = y X 1,..., X n = x) = P(X..., X 1, n = x Y)P(Y = y) P(X 1,..., X n = x)

22 Generic aïve Bayes Model Support: Depends on the choice of event model, P(X Y) Model: Product of prior and the event model P (,Y)=P (Y ) K =1 P (X Y ) Training: Find the class- conditional MLE parameters For P(Y), we find the MLE using all the data. For each P(X Y) we condition on the data with the corresponding class. Classification: Find the class that maximizes the posterior ŷ = y p(y ) 22

23 Generic aïve Bayes Model Classification: ŷ = y = y = y p(y ) p( y)p(y) p(x) p( y)p(y) (posterior) (by Bayes rule) 23

24 Model 1: Bernoulli aïve Bayes Support: Binary vectors of length K {0, 1} K Generative Story: Y Bernoulli( ) X Bernoulli(,Y ) {1,...,K} Model: p, (x,y)=p, (x 1,...,x K,y) = p (y) K =1 p (x y) =( ) y (1 ) (1 y) K =1 (,y ) x (1,y) (1 x ) 24

25 Model 0: ot- so- naïve Model? Flip weighted coin If HEADS, flip each red coin If TAILS, flip each blue coin y x 1 x 2 x 3 x K Each red coin corresponds to an x We can generate data in this fashion. Though in practice we never would since our data is given. Instead, this provides an explanation of how the data was generated (albeit a terrible one). 25

26 Model 1: Bernoulli aïve Bayes Support: Binary vectors of length K {0, 1} K Generative Story: Y Bernoulli( ) X Bernoulli(,Y ) {1,...,K} Model: p, (x,y) =( ) y (1 ) (1 y) K =1 Same as Generic aïve Bayes (,y ) x (1,y) (1 x ) Classification: Find the class that maximizes the posterior ŷ = y p(y ) 26

27 Generic aïve Bayes Model Recall Classification: ŷ = y = y = y p(y ) p( y)p(y) p(x) p( y)p(y) (posterior) (by Bayes rule) 27

28 Model 1: Bernoulli aïve Bayes Training: Find the class- conditional MLE parameters For P(Y), we find the MLE using all the data. For each P(X Y) we condition on the data with the corresponding class. =,0 =,1 = i=1 I(y(i) = 1) i=1 I(y(i) =0 x (i) = 1) i=1 I(y(i) = 0) i=1 I(y(i) =1 x (i) = 1) i=1 I(y(i) = 1) {1,...,K} 28

29 Model 1: Bernoulli aïve Bayes Training: Find the class- conditional MLE parameters For P(Y), we find the MLE using all the data. For each P(X Y) we condition on the data with the corresponding class. =,0 =,1 = i=1 I(y(i) = 1) i=1 I(y(i) =0 x (i) = 1) i=1 I(y(i) = 0) i=1 I(y(i) =1 x (i) = 1) i=1 I(y(i) = 1) Data: y x 1 x 2 x 3 x K {1,...,K}

30 Outline Bacground: Maximum lielihood estimation (MLE) Maximum a posteriori (MAP) estimation Example: Exponential distribution Generative Models Model 0: ot- so- naïve Model aïve Bayes aïve Bayes Assumption Model 1: Bernoulli aïve Bayes Model 2: Multinomial aïve Bayes Model 3: Gaussian aïve Bayes Model 4: Multiclass aïve Bayes Smoothing Add- 1 Smoothing Add- λ Smoothing MAP Estimation (Beta Prior) 30

31 Smoothing 1. Add- 1 Smoothing 2. Add- λ Smoothing 3. MAP Estimation (Beta Prior) 31

32 MLE What does maximizing lielihood accomplish? There is only a finite amount of probability mass (i.e. sum- to- one constraint) MLE tries to allocate as much probability mass as possible to the things we have observed at the expense of the things we have not observed 32

33 MLE For aïve Bayes, suppose we never observe the word serious in an Onion article. In this case, what is the MLE of p(x y)?,0 = i=1 I(y(i) =0 x (i) = 1) i=1 I(y(i) = 0) ow suppose we observe the word serious at test time. What is the posterior probability that the article was an Onion article? p(y x) = p(x y)p(y) p(x) 33

34 1. Add- 1 Smoothing The simplest setting for smoothing simply adds a single pseudo-observation to the data. This converts the true observations D into a new dataset D from we derive the MLEs. D = {( (i),y (i) )} i=1 (1) D = D {(, 0), (, 1), (, 0), (, 1)} (2) where is the vector of all zeros and is the vector of all ones. This has the effect of pretending that we observed each feature x with each class y. 34

35 1. Add- 1 Smoothing What if we write the MLEs in terms of the original dataset D? = i=1 I(y(i) = 1),0 = 1+ i=1 I(y(i) =0 x (i) = 1) 2+ i=1 I(y(i) = 0),1 = 1+ i=1 I(y(i) =1 x (i) = 1) 2+ i=1 I(y(i) = 1) {1,...,K} 35

36 2. Add- λ Smoothing For the Categorical Distribution Suppose we have a dataset obtained by repeatedly rolling a K-sided (weighted) die. Given data D = {x (i) } i=1 where x(i) {1,...,K}, we have the following MLE: = i=1 I(x(i) = ) With add- smoothing, we add pseudo-observations as before to obtain a smoothed estimate: = + i=1 I(x(i) = ) + 36

37 MLE vs. MAP Recall Suppose we have data D = {x (i) } i=1 MLE = MAP = i=1 i=1 p( (i) ) p( (i) )p( ) Maximum Lielihood Estimate (MLE) Maximum a posteriori (MAP) estimate Prior 37

38 3. MAP Estimation (Beta Prior) Generative Story: The parameters are drawn once for the entire dataset. for for i {1,...,K}: for y {0, 1}:,y Beta(, ) {1,...,}: y (i) Bernoulli( ) for x (i) {1,...,K}: Bernoulli(,y (i)) Training: Find the class- conditional MAP parameters = i=1 I(y(i) = 1),0 = ( 1) + i=1 I(y(i) =0 x (i) = 1) ( 1) + ( 1) + i=1 I(y(i) = 0),1 = ( 1) + i=1 I(y(i) =1 x (i) = 1) ( 1) + ( 1) + i=1 I(y(i) = 1) {1,...,K} 38

39 Outline Bacground: Maximum lielihood estimation (MLE) Maximum a posteriori (MAP) estimation Example: Exponential distribution Generative Models Model 0: ot- so- naïve Model aïve Bayes aïve Bayes Assumption Model 1: Bernoulli aïve Bayes Model 2: Multinomial aïve Bayes Model 3: Gaussian aïve Bayes Model 4: Multiclass aïve Bayes Smoothing Add- 1 Smoothing Add- λ Smoothing MAP Estimation (Beta Prior) 39

40 Model 2: Multinomial aïve Bayes Support: Option 1: Integer vector (word IDs) =[x 1,x 2,...,x M ] where x m {1,...,K} a word id. Generative Story: for i {1,...,}: y (i) Bernoulli( ) for j {1,...,M i }: x (i) j Multinomial( y (i), 1) Model: p, (x,y)=p (y) K =1 p (x y) =( ) y (1 ) (1 y) M i j=1 y,x j 40

41 Gaussian Discriminative Analysis l learning f: X Y, where l X is a vector of real-valued features, X n = < X n1, X n m > l Y is an indicator vector l What does that imply about the form of P(Y X)? l The joint probability of a datum and its label is: l Given a datum x n, we predict its label using the conditional probability of the label given the datum: Eric CMU, Y n X n { } ) - ( ) - ( exp - ) (2 1 ), 1, ( 1) ( ), 1, ( / 1 n T n n n n n n y p y p y p µ µ π π µ σ µ!! x x x x Σ Σ = Σ = = = = { } { } Σ Σ Σ Σ = Σ = ' ' / ' / ) - ( ) - ( - exp ) (2 1 ) - ( ) - ( - exp ) (2 1 ),, 1 ( n T n n T n n y n p µ µ π π µ µ π π µ!!!! x x x x x

42 Model 3: Gaussian aïve Bayes Support: Model: Product of prior and the event model R K p(x,y)=p(x 1,...,x K,y) = p(y) K =1 p(x y) Gaussian aive Bayes assumes that p(x y) is given by a ormal distribution. 42

43 Gaussian aïve Bayes Classifier l When X is multivariate-gaussian vector: l The joint probability of a datum and it label is: l The naïve Bayes simplification l More generally: l Where p(..) is an arbitrary conditional (discrete or continuous) 1-D density Eric CMU, { } ) - ( ) - ( exp - ) (2 1 ), 1, ( 1) ( ), 1, ( / 1 n T n n n n n n y p y p y p µ µ π π µ µ!!!! x x x x Σ Σ = Σ = = = Σ = Y n X n = = = = = j j j j n j j j j n j n n n n x y x p y p y p exp 2 1 ), 1, ( 1) ( ), 1, ( σ µ πσ π σ µ σ µ x = = m j n j n n n n y x p y p y p 1 ), ( ) ( ),, ( η π π η x Y n X n,1 X n,2 X n,m

44 VISUALIZIG AÏVE BAYES Slides in this section from William Cohen (10-601B, Spring 2016) 44

45

46 Fisher Iris Dataset Fisher (1936) used 150 measurements of flowers from 3 different species: Iris setosa (0), Iris virginica (1), Iris versicolor (2) collected by Anderson (1936) Species Sepal Length Sepal Width Petal Length Petal Width Full dataset: 46

47 %% Import the IRIS data load fisheriris; X = meas; pos = strcmp(species,'setosa'); Y = 2*pos - 1; %% Visualize the data imagesc([x,y]); title('iris data');

48 %% Visualize by scatter plotting the the first two dimensions figure; scatter(x(y<0,1),x(y<0,2),'r*'); hold on; scatter(x(y>0,1),x(y>0,2),'bo'); title('iris data');

49 %% Compute the mean and SD of each class PosMean = mean(x(y>0,:)); PosSD = std(x(y>0,:)); egmean = mean(x(y<0,:)); egsd = std(x(y<0,:)); %% Compute the B probabilities for each class for each grid element [G1,G2]=meshgrid(3:0.1:8, 2:0.1:5); Z1 = gaussmf(g1,[possd(1),posmean(1)]); Z2 = gaussmf(g2,[possd(2),posmean(2)]); Z = Z1.* Z2; V1 = gaussmf(g1,[egsd(1),egmean(1)]); V2 = gaussmf(g2,[egsd(2),egmean(2)]); V = V1.* V2;

50 % Add them to the scatter plot figure; scatter(x(y<0,1),x(y<0,2),'r*'); hold on; scatter(x(y>0,1),x(y>0,2),'bo'); contour(g1,g2,z); contour(g1,g2,v);

51 %% ow plot the difference of the probabilities figure; scatter(x(y<0,1),x(y<0,2),'r*'); hold on; scatter(x(y>0,1),x(y>0,2),'bo'); contour(g1,g2,z- V); mesh(g1,g2,z- V); alpha(0.4)

52 AÏVE BAYES IS LIEAR Slides in this section from William Cohen (10-601B, Spring 2016)

53 Question: what does the boundary between positive and negative loo lie for aïve Bayes?

54

55 argmax y P(X i = x i Y = y ) P(Y = y ) i = argmax y log P(X i = x i Y = y ) + log P(Y = y ) i = argmax y {+1, 1} log P(x i y ) + log P(y ) # & = sign% log P(x i y +1 ) log P(x i y 1 ) + log P(y +1 ) log P(y 1 )( $ ' i i i two classes only # = sign log P(x i y +1 ) P(x i y 1 ) + log P(y +1) & % ( $ P(y 1 )' i rearrange terms

56 argmax y P(X i = x i Y = y ) P(Y = y ) i # = sign log P(x i y +1 ) P(x i y 1 ) + log P(y +1) & % ( $ P(y 1 )' i if x i = 1 or 0. " u i = log P(x =1 y ) i +1 P(x i =1 y 1 ) log P(x = 0 y ) % i +1 $ ' # P(x i = 0 y 1 )& " " = sign x i log P(x =1 y ) i +1 P(x i =1 y 1 ) log P(x = 0 y ) % " i +1 $ ' + log P(x = 0 y ) % i +1 $ ' + log P(y ) % +1 $ # i # P(x i = 0 y 1 )& i # P(x i = 0 y 1 )& P(y 1 ) ' & " % = sign$ x i u i + u 0 ' # i & " % = sign$ x i u i + x 0 u 0 ' # i & x 0 =1 for every x (bias term) = sign x u ( ) " u 0 = log P(x i = 0 y +1 )% $ ' + log P(y +1) # P(x i = 0 y 1 )& P(y 1 ) i

57

58

59 Why don t we drop the generative model and try to learn this hyperplane directly?

60 Model 4: Multiclass aïve Bayes Model: The only change is that we permit y to range over C classes. p(x,y)=p(x 1,...,x K,y) = p(y) K =1 p(x y) ow, y Multinomial(, 1) and we have a separate conditional distribution p(x y) for each of the C classes. 60

61 aïve Bayes Assumption Recall Conditional independence of features: P (X 1,...,X K,Y)=P(X 1,...,X K Y )P (Y ) = K =1 P (X Y ) P (Y ) 61

62 =P (X 1,...,X K Y )P (Y ) = K =1 P (X Y ) P (Y ) What s wrong with the aïve Bayes Assumption? The features might not be independent!! Example 1: If a document contains the word Donald, it s extremely liely to contain the word Trump These are not independent! Example 2: If the petal width is very high, the petal length is also liely to be very high 62

63 Summary 1. aïve Bayes provides a framewor for generative modeling 2. Choose an event model appropriate to the data 3. Train by MLE or MAP 4. Classify by maximizing the posterior 63