A crash course in probability and Naïve Bayes classification

Transcription

1 Probability theory A crash course in probability and Naïve Bayes classification Chapter 9 Random variable: a variable whose possible values are numerical outcomes of a random phenomenon. s: A person s height, the outcome of a coin toss Distinguish between discrete and continuous variables. The distribution of a discrete random variable: The probabilities of each value it can take. Notation: P(X = x i ). These numbers satisfy: X P (X = x i )=1 i 1 2 Probability theory Probability theory A joint probability distribution for two variables is a table. Marginal Probability If the two variables are binary, how many parameters does it have? Joint Probability Conditional Probability What about joint probability of d variables P(X 1,,X d )? How many parameters does it have if each variable is binary? 4 1

2 Probability theory The Rules of Probability Marginalization: Marginalization Product Rule Product Rule Independence: X and Y are independent if P(Y X) = P(Y) This implies P(X,Y) = P(X) P(Y) Using probability in learning P (Y X) Interested in: 1% X features Y labels 0.5% 0% For example, when classifying spam, we could estimate P(Y Viagara, lottery) We would then classify an example if P(Y X) > 0.5. However, it s usually easier to model P(X Y) Maximum likelihood Fit a probabilistic model P(x θ) to data Estimate θ Given independent identically distributed (i.i.d.) data X = (x 1, x 2,, x n ) Likelihood Log likelihood P (X ) =P (x 1 )P (x 2 ),...,P(x n ) ln P (X ) = nx ln P (x i ) i=1 Maximum likelihood solution: parameters θ that maximize ln P(X θ) 7 2

3 : coin toss Estimate the probability p that a coin lands Heads using the result of n coin tosses, h of which resulted in heads. The likelihood of the data: Log likelihood: Taking a derivative and setting to 0: P (X ) =p h (1 p) n h ln P (X ) =h ln p +(n h)ln(1 ln P (X = h p (n h) (1 p) =0 From the product rule: and: Therefore: P(Y, X) = P(Y X) P(X) P(Y, X) = P(X Y) P(Y) P (Y X) = This is known as Bayes rule Bayes rule P (X Y )P (Y ) P (X) ) p = h n 9 10 P (Y X) = posterior Bayes rule P (X Y )P (Y ) P (X) posterior likelihood prior P(X) can be computed as: P (X) = X Y P (X Y )P (Y ) likelihood prior Maximum a-posteriori and maximum likelihood The maximum a posteriori (MAP) rule: y MAP = arg max Y P (X Y )P (Y ) P (Y X) = arg max Y P (X) If we ignore the prior distribution or assume it is uniform we obtain the maximum likelihood rule: y ML = arg max P (X Y ) Y = arg max P (X Y )P (Y ) Y A classifier that has access to P(Y X) is a Bayes optimal classifier. But is not important for inferring a label 12 3

4 Naïve Bayes classifier We would like to model P(X Y), where X is a feature vector, and Y is its associated label. Task:%Predict%whether%or%not%a%picnic%spot%is%enjoyable% % Training&Data:&& X%=%(X 1 %%%%%%%X 2 %%%%%%%%X 3 %%%%%%%% %%%%%%%% %%%%%%%X d )%%%%%%%%%%Y% Naïve Bayes classifier We would like to model P(X Y), where X is a feature vector, and Y is its associated label. Simplifying assumption: conditional independence: given the class label the features are independent, i.e. P (X Y )=P (x 1 Y )P (x 2 Y ),...,P(x d Y ) n&rows& How many parameters now? How many parameters? Prior: P(Y) k-1 if k classes Likelihood: P(X Y) (2 d 1)k for binary features Naïve Bayes classifier We would like to model P(X Y), where X is a feature vector, and Y is its associated label. Simplifying assumption: conditional independence: given the class label the features are independent, i.e. P (X Y )=P (x 1 Y )P (x 2 Y ),...,P(x d Y ) Naïve Bayes decision rule: y NB = arg max Y Naïve Bayes classifier P (X Y )P (Y ) = arg max Y dy P (x i Y )P (Y ) i=1 If conditional independence holds, NB is an optimal classifier! How many parameters now? dk + k

5 Training a Naïve Bayes classifier Training data: Feature matrix X (n x d) and labels y 1, y n Maximum likelihood estimates: Class prior: Likelihood: ˆP (y) = {i : y i = y} n ˆP (x i y) = ˆP (x i,y) ˆP (y) = {i : X ij = x i,y i = y} /n {i : y i = y} /n classification Suppose our vocabulary contains three words a, b and c, and we use a multivariate Bernoulli model for our s, with parameters = (0.5,0.67,0.33) = (0.67,0.33,0.33) This means, for example, that the presence of b is twice as likely in spam (+), compared with ham. The to be classified contains words a and b but not c, and hence is described by the bit vector x = (1,1,0). We obtain likelihoods P(x ) = (1 0.33) = P(x ) = (1 0.33) = The ML classification of x is thus spam classification: training data a? b? c? Class e e e e e e e e ount vectors. (right) The same data set What are the parameters of the model? classification: training data a? b? c? Class e e e e e e e e ount vectors. (right) The same data set What are the parameters of the model? ˆP (y) = {i : y i = y} n ˆP (x i y) = ˆP (x i,y) ˆP (y) = {i : X ij = x i,y i = y} /n {i : y i = y} /n

6 classification: training data a? b? c? Class e e e e e e e e ount vectors. (right) The same data set What are the parameters of the model? P(+) = 0.5, P(-) = 0.5 P(a +) = 0.5, P(a -) = 0.75 P(b +) = 0.75, P(b -)= 0.25 P(c +) = 0.25, P(c -)= 0.25 ˆP (x i y) = ˆP (x i,y) ˆP (y) ˆP (y) = {i : y i = y} n = {i : Xij = xi,yi = y} /n {i : y i = y} /n Comments on Naïve Bayes Usually features are not conditionally independent, i.e. P (X Y ) 6= P (x 1 Y )P (x 2 Y ),...,P(x d Y ) And yet, one of the most widely used classifiers. Easy to train! It often performs well even when the assumption is violated. Domingos, P., & Pazzani, M. (1997). Beyond Independence: Conditions for the Optimality of the Simple Bayesian Classifier. Machine Learning. 29, When there are few training examples What if you never see a training example where x 1 =a when y=spam? P(x spam) = P(a spam) P(b spam) P(c spam) = 0 What to do? When there are few training examples What if you never see a training example where x 1 =a when y=spam? P(x spam) = P(a spam) P(b spam) P(c spam) = 0 What to do? Add virtual examples for which x 1 =a when y=spam

7 Naïve Bayes for continuous variables Continuous Probability Distributions Need to talk about continuous distributions! The probability of the random variable assuming a value within some given interval from x 1 to x 2 is defined to be the area under the graph of the probability density function between x 1 and x 2. f (x) Uniform f (x) Normal/Gaussian x x x x 1 x 1 x Expectations The Gaussian (normal) distribution Discrete variables Continuous variables Conditional expectation (discrete) Approximate expectation (discrete and continuous) 7

8 Properties of the Gaussian distribution Standard Normal Distribution 99.72% 95.44% 68.26% A random variable having a normal distribution with a mean of 0 and a standard deviation of 1 is said to have a standard normal probability distribution. µ 3σ µ 1σ µ 2σ µ µ + 1σ µ + 2σ µ + 3σ x Standard Normal Probability Distribution Gaussian Parameter Estimation Converting to the Standard Normal Distribution z x = µ σ Likelihood function We can think of z as a measure of the number of standard deviations x is from µ. 8

9 Maximum (Log) Likelihood 34 Gaussian models Assume we have data that belongs to three classes, and assume a likelihood that follows a Gaussian distribution Likelihood function: P (X i = x Y = y k )= Gaussian Naïve Bayes 1 (x µik ) 2 p exp 2 ik 2 2 ik Need to estimate mean and variance for each feature in each class

10 Summary Naïve Bayes classifier: ² ² ² What s the assumption Why we make it How we learn it Naïve Bayes for discrete data Gaussian naïve Bayes 37 10