Discrimination: finding the features that separate known groups in a multivariate sample.

Transcription

1 Discrimination and Classification Goals: Discrimination: finding the features that separate known groups in a multivariate sample. Classification: developing a rule to allocate a new object into one of a number of known groups. A classification rule is based on the features that separate the groups, so the goals overlap. 1

2 Simplest Case: Two Groups Both discrimination and classification depend on multivariate observation X. Basic setup: Consider each group a population: π 1 and π 2. The density of X in π g is f g (x), g = 1, 2. 2

3 Alternatively: A single population with a group indicator G. The density of X given G = g is f g (x). P(G = 1) = p 1, P(G = 2) = p 2 = 1 p 1. p 1 and p 2 are the prior probabilities of the groups. 3

4 Example: ownership of a riding mower. Lotsize Owner Non owner Income 4

5 Classification problem is to predict whether a homeowner owns a riding mower, given the Income and Lot Size. Rule must give an answer for any x, so it defines two regions R 1 and R 2. Classification errors: P (2 1) = P(X R 2 G = 1) = P (1 2) = P(X R 1 G = 2) = R 2 f 1 (x)dx, R 1 f 2 (x)dx. 5

6 Unconditionally: P (correctly classified as π 1 ) = P (1 1)p 1 P (missclassified as π 1 ) = P (1 2)p 2 P (correctly classified as π 2 ) = P (2 2)p 2 P (missclassified as π 2 ) = P (2 1)p 1 The two types of classification error may have different costs: Classify as: π 1 π 2 True population: π 1 0 c(2 1) π 2 c(1 2) 0 6

7 Expected Cost of Misclassification: ECM = c(2 1)P (2 1)p 1 + c(1 2)P (1 2)p 2. The classification region that minimizes ECM is ( ) ( ) f R 1 : 1 (x) c(1 2) p2 f 2 (x). c(2 1) p 1 Note: these regions depend on the ratios of: the densities; the costs of misclassification; the prior probabilities. 7

8 Special cases: Total Probability of Misclassification: TPM = P(misclassify either way) = P (2 1)p 1 + P (1 2)p 2 is ECM with c(1 2) = c(2 1) = 1. Posterior mode: allocate to the population with the higher posterior probability; by Bayes s rule: P (π 1 x 0 ) = P (π 2 x 0 ) = same rule as minimizing TPM. p 1 f 1 (x 0 ) p 1 f 1 (x 0 ) + p 2 f 2 (x 0 ) p 2 f 2 (x 0 ) p 1 f 1 (x 0 ) + p 2 f 2 (x 0 ) 8

9 Multivariate Normal Populations If f 1 (x) and f 2 (x) are multivariate normal with the same Σ and means µ 1 and µ 2, then the minimum-ecm allocation rule is based on a linear function of x. Basic result is [ ] [ f1 (x) ln = (µ f 2 (x) 1 µ 2 ) Σ 1 x 1 ] 2 (µ 1 + µ 2) Optimal region is R 1 : (µ 1 µ 2 ) Σ [x 1 1 ] 2 (µ 1 + µ 2) ln [( c(1 2) c(2 1) ) ( )] p2 p 1. 9

10 Typically, Σ, µ 1 and µ 2 are unknown, but replaced with sample estimates, based on training data. Example: Hemophilia A carriers. Women subjects in two groups: Non-carriers and Obligatory carriers. Variables: X 1 = log 10 (AHF acitivity) X 2 = log 10 (AHF-like antigen) SAS proc discrim program and output. 10

11 Evaluating Classification Functions The Optimum Error Rate is the minimum TPM. For multivariate normal populations, OER = 1 Φ ( ) where 2 is the Generalized Squared Distance between groups: 2 = (µ 1 µ 2 ) Σ 1 (µ 1 µ 2 ). 2, For the hemophilia example, SAS reports 2 = , whence OER =

12 The OER calculation is the estimated error rate for allocation using the true parameters good in large samples, but dubious in small samples. Sample-based evaluation: APparent Error Rate = fraction of training data that are misclassified; reported by SAS as Resubstitution Summary ; APER = for the example; under-estimates error rate, because the data used to develop the rule are then used to evaluate it; Lachenbruch s holdout procedure: omit one observation at a time, recalculate the classification rule, and classify the omitted observation; reported by SAS as Crossvalidation Summary, for the example. 12

13 Quadratic Classification If f 1 (x) and f 2 (x) are multivariate normal with different covariances Σ 1 and Σ 2 and means µ 1 and µ 2, then the minimum- ECM allocation rule is based on a quadratic function of x. Classification regions are now bounded by conic sections (generally elliptical or hyperbolic). No simple explicit representation. 13

14 In proc discrim, use pool = no on the proc statement. For the AHF example, the resubstition Total error rate is and the cross-validated Total error rate is The cross-validated Total error rate is higher than for linear classification, suggesting that the estimation cost of the additional parameters in the two Σ s outweighs their value in improving the fit. 14

15 Nonparametric Classification One alternative to the multivariate normal assumption is classification based on nonparametric density estimates of f g (x). In SAS, use method = npar on the proc discrim statement. You have the choice of nearest neighbor and kernel estimation: for nearest neighbor estimation, use k =... to specify the number of neighbors; for kernel estimation, use kernel =... to name a kernel, and r =... to specify the radius (bandwidth). 15