This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: 1 Bayes Optimality of WaveletBased Discrimination Woojin Chang, SeongHee Kim, and Brani Vidakovic Seoul University and Georgia Institute of Technology ISBA 2004 Vi ˜ na del Mar, Chile May 25, 2004 2 Overview Talk about classifying Y into one of two classes labeled by or 1 , by taking into account predictor X . Definitions and Notation. Bayes Discriminators WaveletBased Approximation Bayes Optimality (or IL 2Consistency) of the Waveletbased Classifier. Simulations and Paper Production Example 3 Definitions ( X,Y ) ∈ IR d ×{ , 1 } . μ ( A ) = P ( X ∈ A ) , A ∈ B ; η ( x ) = P ( Y = 1  X = x ) = E ( Y  X = x ) . Pair ( μ,η ) uniquely determines joint distribution of ( X,Y ) . Any function g : IR d → { , 1 } is a classifier. Bayes Classifier: g * ( x ) = 1 ( η ( x ) > 1 / 2). L ( g ) = P ( g ( X ) 6 = Y ) . [Error, Risk, Misclassification Probability] Result: ( ∀ g ) L ( g * ) ≤ L ( g ) . L * = L ( g * ) Bayes Error [Risk, Probability]. 4 Definitions, contd Assume density of X exists, X ∼ f. Let f and f 1 be classconditional densities, i.e., densities for X  Y = 0 and X  Y = 1 . Let π = P ( Y = 1) and 1 π = P ( Y = 0) be classprobabilities. Function α ( x ) = πf 1 ( x ) (1 π ) f ( x ) has representation (2 η ( x ) 1) f ( x ) . Bayes Classifier: g * ( x ) = 1 ( α ( x ) > 0). • L * = 1 / 2 1 / 2 E (  2 η ( X ) 1  ) • L * = R ((1 π ) f ∧ πf 1 ) dx • π = 1 / 2 , L * = 1 / 2 1 / 4 R  f ( x ) f 1 ( x )  dx 5 Definitions, contd D n = { ( X 1 ,Y 1 ) ,..., ( X n ,Y n ) } training set . Let X be a new observation. g n ( X ) = g n ( X,D n ), a sequence of classification rules. L n = P ( Y 6 = g n ( X,D n )  D n ) . IE L n = P ( Y 6 = g n ( X )) determined by distribution ( X,Y ) and classifier g n . Classifier g n is consistent (weakly): lim n →∞ IE L n = L * ....
View
Full
Document
This note was uploaded on 10/23/2011 for the course ISYE 8843 taught by Professor Vidakovic during the Spring '11 term at Georgia Tech.
 Spring '11
 VIDAKOVIC

Click to edit the document details