Lecture 5: LDA and Logistic Regression

Transcription

1 Lecture 5: and Logistic Regression Hao Helen Zhang Hao Helen Zhang Lecture 5: and Logistic Regression 1 / 39

2 Outline Linear Classification Methods Two Popular Linear Models for Classification Linear Discriminant Analysis () Logistic Regression Models Take-home message: Both and Logistic regression models rely on the linear-odd assumption, indirectly or directly. However, they estimate the coefficients in a different manner. Hao Helen Zhang Lecture 5: and Logistic Regression 2 / 39

3 Linear Classifier Linear Classification Methods Linear methods: The decision boundary is linear. Common linear classification methods: Linear regression methods (covered in Lecture 3) Linear log-odds (logit) models Linear logistic models Linear discriminant analysis () separating hyperplanes (introduced later) perceptron model (Rosenblatt 1958) Optimal separating hyperplane (Vapnik 1996) SVMs From now on, we assume equal costs (by default). Hao Helen Zhang Lecture 5: and Logistic Regression 3 / 39

4 Odds, Logit, and Linear Some terminologies Call the term Call log Pr(Y =1 X=x) Pr(Y =0 X=x) Pr(Y =1 X=x) Pr(Y =0 X=x) is called odds log of the odds, or logit function Linear odds models assume: the logit is linear in x, i.e., log Examples:, Logistic regression Pr(Y = 1 X = x) Pr(Y = 0 X = x) = β 0 + β T 1 x. Hao Helen Zhang Lecture 5: and Logistic Regression 4 / 39

5 Classifier Based on Linear Odd Models From the linear odds, we can obtain posterior class probabilities Pr(Y = 1 x) = Pr(Y = 0 x) = exp (β 0 + β1 T x) 1 + exp (β 0 + β1 T x) exp (β 0 + β1 T x) Assuming equal costs, the decision boundary is given by {x P(Y = 1 x) = 0.5} = {x β 0 + β T 1 x = 0}, which can be interpreted as zero log-odds Hao Helen Zhang Lecture 5: and Logistic Regression 5 / 39

6 logit function log[s/(1 s)] logistic curve exp(t)/(1+exp(t)) g(s) p(t) s t Hao Helen Zhang Lecture 5: and Logistic Regression 6 / 39

7 Linear Discriminant Analysis () assumes Assume each class density is multivariate Gaussian, i.e., X Y j N(µ j, Σ j ), j = 0, 1. Equal covariance assumption Σ j = Σ, j = 0, 1. In other words, both classes are from Gaussian and they have the same covariance matrix. Hao Helen Zhang Lecture 5: and Logistic Regression 7 / 39

8 Linear Discriminant Function Under the mixture Gaussian assumption, the log-odd is log Pr(Y = 1 X = x) Pr(Y = 0 X = x) = log π 1 π (µ 1 + µ 0 ) T Σ 1 (µ 1 µ 0 ) + x T Σ 1 (µ 1 µ 0 ) Under equal costs, the classifies to 1 if and only if [ log π 1 1 ] π 0 2 (µ 1 + µ 0 ) T Σ 1 (µ 1 µ 0 ) + x T Σ 1 (µ 1 µ 0 ) > 0. It has a linear boundary {x : β 0 + x T β 1 = 0}, with β 0 = log π 1 1 π 0 2 (µ 1 + µ 0 ) T Σ 1 (µ 1 µ 0 ), β 1 = Σ 1 (µ 1 µ 0 ). Hao Helen Zhang Lecture 5: and Logistic Regression 8 / 39

9 Parameter Estimation in In practice, π 1, π 0, µ 0, µ 1, Σ are unknown We estimate the parameters from the training data, using MLE or the moment estimator ˆπ j = n j /n, where n k is the size size of class j. ˆµ j = Y i =j x i/n j for j = 0, 1. The sample covariance matrix is S j = 1 n j 1 (x i ˆµ j )(x i ˆµ j ) T Y i =j (Unbiased) pooled sample covariance is a weighted average ˆΣ = = n 0 1 (n 0 1) + (n 1 1) S n (n 0 1) + (n 1 1) S 1 1 (x i ˆµ j )(x i ˆµ j ) T /(n 2) j=0 Y i =j Hao Helen Zhang Lecture 5: and Logistic Regression 9 / 39

10 R code for Fitting (I) There are two ways to call the function lda. The first way is to use a formula and an optional data frame. library(mass) lda(formula, data,subset) Arguments: Output: formula: the form groups x 1 + x , where the response is the grouping factor and the right hand side specifies the (non-factor) discriminators. data: data frame from which variables specified subset: An index vector specifying the cases to be used in the training sample. an object of class lda with multiple components Hao Helen Zhang Lecture 5: and Logistic Regression 10 / 39

11 R code for Fitting (II) The second way is to use a matrix and group factor as the first two arguments. library(mass) lda(x, grouping, prior = proportions, CV = FALSE) Arguments: Output: x: a matrix or data frame or Matrix containing predictors. grouping: a factor specifying the class for each observation. prior: the prior probabilities of class membership. If unspecified, the class proportions for the training set are used. If CV = TRUE, the return value is a list with components class (the MAP classification, a factor) and posterior (posterior probabilities for the classes). Hao Helen Zhang Lecture 5: and Logistic Regression 11 / 39

12 R code for Prediction We use the predict or predict.lda function to classify multivariate observations with lda predict(object, newdata,...) Arguments: Output: object: object of class lda newdata: data frame of cases to be classified or, if object has a formula, a data frame with columns of the same names as the variables used. a list with the components class (the MAP classification, a factor) and posterior (posterior probabilities for the classes) Hao Helen Zhang Lecture 5: and Logistic Regression 12 / 39

13 Fisher s Iris Data (Three-Classification Problems) Fisher (1936) The use of multiple measurements in taxonomic problems. Three species: Iris setosa, Iris versicolor, Iris verginica Four features: the length and the width of the sepals and petals, in centimeters. 50 samples from each species In the following analysis, we randomly select 50% of the data points as the training set, and the rest as the test set. Hao Helen Zhang Lecture 5: and Logistic Regression 13 / 39

14 Hao Helen Zhang Lecture 5: and Logistic Regression 14 / 39

15 Illustration 1 Linear Classification Methods Iris <- data.frame(rbind(iris3[,,1], iris3[,,2], iris3[,,3]), Sp = rep(c("s","c","v"), rep(50,3))) train <- sample(1:150, 75) table(iris$sp[train]) z <- lda(sp ~., Iris, prior = c(1,1,1)/3, subset = train) Hao Helen Zhang Lecture 5: and Logistic Regression 15 / 39

16 Training Error and Test Error #training error rate ytrain <- predict(z, Iris[train, ])$class table(ytrain, Iris$Sp[train]) train_err <- mean(ytrain!=iris$sp[train]) #test error rate ytest <- predict(z, Iris[-train, ])$class table(ytest, Iris$Sp[-train]) test_err <- mean(ytest!=iris$sp[-train]) Hao Helen Zhang Lecture 5: and Logistic Regression 16 / 39

17 Illustration 2 Linear Classification Methods tr <- sample(1:50, 25) train <- rbind(iris3[tr,,1], iris3[tr,,2], iris3[tr,,3]) test <- rbind(iris3[-tr,,1], iris3[-tr,,2], iris3[-tr,,3]) cl <- factor(c(rep("s",25), rep("c",25), rep("v",25))) z <- lda(train, cl) ytest <- predict(z, test)$class Hao Helen Zhang Lecture 5: and Logistic Regression 17 / 39

18 Logistic Regression Linear Classification Methods Model assumption: the log-odd is linear in x. Define log Pr(Y = 1 X = x) Pr(Y = 0 X = x) = β 0 + β T 1 x. p(x; β) = exp (β 0 + β T 1 x) 1 + exp (β 0 + β T l x) Write β = (β 0, β 1 ). The posterior class probabilities can be calculated as p 1 (x) = Pr(Y = 1 x) = p(x; β), p 0 (x) = Pr(Y = 0 x) = 1 p(x; β). The classification boundary is given by: {x : β 0 + β T 1 x = 0}. Hao Helen Zhang Lecture 5: and Logistic Regression 18 / 39

19 Model Interpretation Denote µ = E(Y X) = P(Y = 1 X) By assuming g(µ) = log[µ/(1 µ)] = β 0 + β1 T X, the logit g connects µ with the linear predictor β 0 + β1 T X. We call g the link function Var(Y X) = µ(1 µ). Hao Helen Zhang Lecture 5: and Logistic Regression 19 / 39

20 Interpretation of β j In logistic regression, odds Xj = odds(..., X j = x + 1,...) odds(..., X j = x,...) = e β j. If X j = 0 or 1, then odds for group with X j = 1 are e β j higher than for group with X j = 0, with other parameters fixed. For rare diseases, when incidence is rare, Pr(Y = 0) 1, then odds Pr(Y = 1) e β j = odds Xj Pr(..., X j = x + 1,...). Pr(..., X j = x,...) in a cancer study, an log-or of 5 means that smokers are e times more likely to develop the cancer Hao Helen Zhang Lecture 5: and Logistic Regression 20 / 39

21 Maximum Likelihood Estimate (MLE) for Logistic Models The joint conditional likelihood of y i given x i is l(β) = n log p yi (x; β), i=1 where In details, l(β) = =. p y (β; x) = p(x; β) y [1 p(x; β)] 1 y. n {y i log p(x i ; β) + (1 y i ) log[1 p(x i, β)]} i=1 n {y i (β 10 + β T 1 x i ) log[1 + exp (β 10 + β T 1 x i )]. i=1 Hao Helen Zhang Lecture 5: and Logistic Regression 21 / 39

22 Score Equations Linear Classification Methods For simplicity, now assume x i has 1 in its first component. l(β) n β = x i [y i p(x i ; β)] = 0, i=1 Totally (d+1) nonlinear equations. The first equation n y i = i=1 n p(x i ; β), i=1 expected number of 1 s = observed number in sample. y = [y 1,, y n ] T p = [p(x 1 ; β old ),, p(x n ; β old )] T. { W = diag p(x i ; β old )(1 p(x i ; β old } )). Hao Helen Zhang Lecture 5: and Logistic Regression 22 / 39

23 Newton-Raphson Algorithm The second-derivative (Hessian) matrix 2 l(β) β β T = n x i x T i p(x i ; β)[1 p(x i ; β)]. i=1 1 Choose an initial value β 0 2 Update β by β new = β old Using matrix notations, we have [ 2 ] 1 l(β) l(β) β β T β old β β old l(β) β = X T (y p), 2 l(β) β β T = X T WX. Hao Helen Zhang Lecture 5: and Logistic Regression 23 / 39

24 Iteratively Re-weighted Least Squares (IRLS) Newton-Raphson step β new = β old + (X T WX ) 1 X T (y p) ( = (X T WX ) 1 X T W X β old ) + W 1 (y p) = (X T WX ) 1 X T W z, where we defined the adjusted response z = X β old + W 1 (y p) Repeatedly solve weighted least squares till convergence. β new = arg min β (z X β)t W (z X β), Weight W, response z, and p change in each iteration. The algorithm can be generalized to K 3 case. Hao Helen Zhang Lecture 5: and Logistic Regression 24 / 39

25 Quadratic Approximations β satisfies a self-consistency relationship: it solves a weighted least square fit with response z i = x T i β + (y i ˆp i ) ˆp i (1 ˆp i ) and the weight w i = ˆp i (1 ˆp i ). β = 0 seems to be a good starting value Typically the algorithm converges, but it is never guaranteed. The weighted residual sum-of-squared is Pearson chi-square statistic n (y i ˆp i ) 2 ˆp i (1 ˆp i ), i=1 a quadratic approximation to the deviance. Hao Helen Zhang Lecture 5: and Logistic Regression 25 / 39

26 Statistical Inferences Using the weighted least squares formulation, we have Asymptotic likelihood theory says: if the model is correct, β is consistent Using central limit theorem, the distribution of β converges to N(β, (X T WX ) 1 ) Model building is costly due to iterations, popular shortcuts: For inclusion of a term, use Rao score test. For exclusion of a term, use Wald test. Neither of these two algorithms require iterative fitting, and are based on the maximum likelihood fit of the current model. Hao Helen Zhang Lecture 5: and Logistic Regression 26 / 39

27 R Code for Logistic Regression logist <- glm(formula, family, data, subset,...) predict(logist) Arguments: Output: formula: an object of class formula family: a description of the error distribution and link function to be used in the model. data: an optional data frame, list or environment containing the variables in the model. subset: an optional vector specifying a subset of observations to be used in the fitting process. returns an object of class inheriting from glm and lm Hao Helen Zhang Lecture 5: and Logistic Regression 27 / 39

28 R Code for Logistic Prediction logist <- glm(y~x, data, family=binomial(link="logit")) predict(logist, newdata) summary(logist) anova(logist) The function predict gives the predicted values, newdata is a data frame The function summary is used to obtain or print a summary of the results The function anova produces an analysis of variance table. Hao Helen Zhang Lecture 5: and Logistic Regression 28 / 39

29 Relationship between and Least Squares (LS) For two-class problems, both and least squares fit a linear boundary β 0 + β T x. Their solutions have the following relationship: The least square regression coefficient ˆβ is proportional to the direction, i.e., ˆβ ˆΣ 1 (ˆµ 1 ˆµ 0 ), where ˆΣ is the pooled sample covariance matrix, and ˆµ k is the sample mean of points from class k, k = 0, 1. In other words, their slope coefficients are identical, up to a scalar multiple. (Exercise 4.2 in textbook) The LS intercept ˆβ 0 is generally different from that of, unless n 1 = n 0. In general, they have different decision rules (unless n 1 = n 2 ). Hao Helen Zhang Lecture 5: and Logistic Regression 29 / 39

30 Common Feature of and Logistic Regression Models For both and logistic regression, the logit has a linear form log Pr(Y = 1 x) Pr(Y = 0 x) = β 0 + β T 1 x. Or equivalently, for both estimators, their posterior class probability can be expressed in the form of Pr(Y = 1 x) = exp (β 0 + β T 1 x) 1 + exp (β 0 + β T 1 x). They have exactly same forms. Are they same estimators? Hao Helen Zhang Lecture 5: and Logistic Regression 30 / 39

31 Major Differences of and Logistsics Regresion Main differences of two estimators include Where is the linear-logit from? What assumptions are made on the data distribution? (Difference on data distribution) How to estimate the linear coefficients? (Difference in parameter estimation) Any assumption on the marginal density of X? (flexibility of model) Hao Helen Zhang Lecture 5: and Logistic Regression 31 / 39

32 Where is Linear-logit from? For, the linear logit is due to the equal-covariance Gaussian assumption on data log Pr(Y = 1 x) Pr(Y = 0 x) = log π 1 π (µ 1 + µ 0 ) T Σ 1 (µ 1 + µ 0 ) +x T Σ 1 (µ 1 µ 0 ) = β 0 + β T 1 x For Logistic model, the linear logit is due to construction log Pr(Y = 1 x) Pr(Y = 0 x) = β 0 + β T 1 x Hao Helen Zhang Lecture 5: and Logistic Regression 32 / 39

33 Difference in Marginal Density Assumption The assumptions on Pr(X): The logistic model leaves the marginal density of X arbitrary and unspecified. The model assumes a Gaussian density Pr(X) = 1 π j φ(x; µ j, Σ) j=0 Conclusion: The logistic model makes less assumptions about the data, and hence is more general. Hao Helen Zhang Lecture 5: and Logistic Regression 33 / 39

34 Difference in Parameter Estimation Logistic regression Maximizing the conditional likelihood, the multinomial likelihood with probabilities Pr(Y = k X) The marginal density Pr(X) is totally ignored (fully nonparametric using the empirical distribution function which places 1/n at each observation) Maximizing the full log-likelihood based on the joint density Pr(X, Y = j) = φ(x; µ j, Σ)π j, Standard MLE theory leads to estimators ˆµ j, ˆΣ, ˆπ j Marginal density does play a role Hao Helen Zhang Lecture 5: and Logistic Regression 34 / 39

35 More Comments Linear Classification Methods is easier to compute than logistic regression. If the true f k (x) s are Gaussian, is better. Logistic regression may lose efficiency around 30% asymptotically in error rate (by Efron 1975) Robustness? uses all the points to estimate the covariance matrix; more information but not robust against outliers Logistic regression down-weights points far from decision boundary (recall the weight is p i (1 p i ); more robust and safer In practice, these two methods often give similar results (for approximately normal distributed data) Hao Helen Zhang Lecture 5: and Logistic Regression 35 / 39

36 Two-dimensional Linear Example Consider the following scenarios for a two-class problem: π 1 = π 0 = 0.5. Class 1: X N 2 ((1, 1) T, 4I) Class 2: X N 2 (( 1, 1) T, 4I) The Bayes boundary is We generate {x : x 1 x 2 = 0}. the training set of n = 200, 400 to fit the classifier the testing set of n = 2000 to evaluate its prediction performance. Hao Helen Zhang Lecture 5: and Logistic Regression 36 / 39

37 Training Scatter Plot x1 x Bayes and Linear Classifiers x1 x2 bay lin Logistic and Classifiers x1 x2 log lda All Linear Classifiers x1 x2 bay lin log lda Hao Helen Zhang Lecture 5: and Logistic Regression 37 / 39

38 Performance of Various Classifiers (n=200) Bayes Linear Logistic Train Error Test Error Fitted classification boundaries: Linear & : x 2 = x 1 Logistic: x 2 = x 1. Hao Helen Zhang Lecture 5: and Logistic Regression 38 / 39

39 Training Scatter Plot x1 x Bayes and Linear Classifiers x1 x2 bay lin Logistic and Classifiers x1 x2 log lda All Linear Classifiers x1 x2 bay lin log lda Hao Helen Zhang Lecture 5: and Logistic Regression 39 / 39