Logistic Regression. May 28, Decision boundary is a property of the hypothesis and not the data set e z. g(z) = g(z) 0.

Transcription

1 Logistic Regression May 28, 202 Logistic Regression. Decision Boundary Decision boundary is a property of the hypothesis and not the data set. sigmoid function: h (x) = g( x) = P (y = x; ) suppose predict y = ifh (x) 0.5 g(z) = + e z g(z) 0.5 when z 0 suppose predict y = 0if h (x) < 0.5 h (x) = g( x) 0.5 whenever x 0 h (x) = g( x) < 0.5 whenever x < 0 Non Linear Decision Boundaries Add higher order polynomials, 2 exra features. h (x) = g( 0 + x + 2 x x x 2 2) 0 = 0 predict y = if + x 2 + x 2 2 0, that is x 2 + x 2 2 if we plot x 2 + x 2 2 =, it s a circle. Inside of the circle, y = 0, outside y =. With even higher polynomial terms, we can get even more complex decision boundaries.

2 .2 Cost Function Training set: m examples: {(x (), y () ), (x (2), y (2) ),..., (x (m), y (m) ) x x 0 x... x n Rn+ x 0 =, y {0, } h (x) = + e x How do we choose parameters? Cost function for logistic regression, needs to be convex and a cost function for linear regression would not be convex. Instead, use (6:20): { log(h (x)) if y = Cost(h (x), y) = log( h (x)) if y = 0 Cost = 0 if y =, h (x) = But as h (x) 0, Cost Captures intuition that if h (x) = 0, (predict P (y = x; ), but y =, we ll penalize learning algorithm by a very large cost..3 Simplified cost function and gradient descent Logistic regression cost function J() = m Cost(h (x (i), y (i) ) m { log(h (x)) if y = Cost(h (x), y) = log( h (x)) if y = 0 Note: y = 0 or always We can rewrite the cost function: Cost(h (x), y) = y log(h (x)) ( y) log( h (x)) So then if y = or if y = 0, this function loses terms accordingly. Now we can rewrite the cost function like so: [ J() = m ] y (i) log h (x (i) ) + ( y (i) ) log( h (x (i) )) m To fit parameters : 2

3 To make a prediction given new x: Output h (x) = min J() + e x which we interpret as probability that y = p(y = x; ) We re going to minimize the cost function with Gradient Descent. Vectorized: min J() : Repeat { } m j = j α (h (x (i) ) y (i) )x (i) j = 0. n m := α (h (x (i) ) y (i) )x (i) This is not the same as Gradient Descent for linear regression, because the definition of h (x)has changed. Apply the method from linear regression to make sure logistic regression is converging, plot cost. Note: Recall: Vectorized: [ J() = m ] y (i) log h (x (i) ) + ( y (i) ) log( h (x (i) )) m J() = m [ y log h (X) ( y) log( h (X)) ] The gradient of the cost is a vector of the same length as where the j th element for j = (0,, 2,..., n) is defined as follows: 3

4 Vectorized: J() j = m J() j m (h (x (i) ) y (i) )x (i) j = m X h (X) y.4 Advanced Optimization Given, we have code that can compute: J() j J() (for j = 0,,..., n) We can use these to minimize the cost function: Gradient descent Conjugate gradient BFGS L-BFGS Advantages: No need to manually pick α Line search algorithm automatically picks α, even for each iteration. Often faster than gradient descent Disadvantages: More complex Example: [ ] = 2 = 5 2 = 5 J() = ( 5) 2 + ( 2 5) 2 J() = 2( 5) 2 J() = 2( 5) function [ jval, gradient ] = costfunction ( theta ) 2 jval = ( theta () - 5) ^2 + ( theta (2) - 5) ^2; 3 gradient = zeros (2, ) 4 gradient () = 2 * ( theta () - 5); 5 gradient () = 2 * ( theta (2) - 5); 4

5 jval gets the result of the cost function gradient gets a 2x vector, the terms of which correspond to the two derivative terms. We can then call the advanced optimization function, fminunc, function minimization unconstrained. Chooses optimal value of automagically. options = optimset ( GradObj, on, MaxIter, 00 ); 2 initialtheta = zeros (2, ); 3 [ opttheta, functionval, exitflag ] =... 4 fminunc initialtheta, options represents a pointer to the above costfunction set some options: GradObj, on means that we will provide a gradient MaxIter, 00 means the number of iterations initialt heta R d, d 2 How do we apply this to logistic regression? 0 theta =. n function [jval, gradient] = costfunction(theta) jval = [code to compute J()]; gradient() = [code to compute 0 J()]; gradient(2) = [code to compute J()];. gradient(n+) = [code to compute n J()];.5 Multi-class classification: One-vs-all How to get logistic regression to get to work with multi-class classification problems. Scenario : tag s for different folders Scenario 2: Medical diagrams: Not ill, cold, flu Scenario 3: Weather: Sunny, Cludy, Rainy, Snow There will be more groups of items on plot. Turn the problem into 3 different binary classification problems. We ll create a fake trainig set, where Class gets assigned to a positive class and Class 2 and 3 get assigned to the negative class. h () h (2) h (2) (x)will denote, say triangles (x)will denote, say squares (x)will denote, say x s 5

6 We fit 3 classifiers that try to estimate what is the probability that y = i h (i) = P (y = i x; ) (i =, 2, 3) Train a logistic regression classifier h (i) x for each class i to predict the probability that y =. On a new input x, to make a prediction, pick the class i that maximizes max h (i) i (x) 6