1

Model Assessment and Selection

Lecture Notes for Comp540 Chapter7

Jian Li

Mar.2007

2

Goal

• Model Selection

• Model Assessment

3

A Regression Problem

• y = f(x) + noise

• Can we learn f from this data?

• Let’s consider three methods...

4

Linear Regression

5

Quadratic Regression

6

Joining the dots

7

Which is best?

• Why not choose the method with the best fit to the data?

“How well are you going to predict future data drawn

from the same distribution?”

8

Model Selection and Assessment

• Model Selection: Estimating performances of different models to choose the best one (produces the minimum of the test error)

• Model Assessment: Having chosen a model, estimating the prediction error on new data

9

Why Errors

• Why do we want to study errors?

• In a data-rich situation split the data:

Train Validation Test

Model Selection Model assessment

• But, that’s not usually the case

10

Overall Motivation

• Errors Measurement of errors (Loss functions) Decomposing Test Error into Bias & Variance

• Estimating the true error Estimating in-sample error (analytically )

AIC, BIC, MDL, SRM with VC Estimating extra-sample error (efficient sample reuse)

Cross Validation & Bootstrapping

11

Measuring Errors: Loss Functions

• Typical regression loss functions Squared error:

Absolute error:

12

Measuring Errors: Loss Functions

• Typical classification loss functions 0-1 Loss:

Log-likelihood (cross-entropy loss / deviance):

13

• We want to minimize generalization error or test error:

The Goal: Low Test Error

Err E[L(Y, ˆ f (X))]

err 1N

L(y i, ˆ f (x i))i1

N

• But all we really know is training error:

• And this is a bad estimate of test error

?

14

Bias, Variance & Complexity

Training error can always be reduced when increasing model complexity, but risks over-fitting.

Typically

15

Decomposing Test Error

For squared-error loss & additive noise:

Irreducible error of target Y

Deviation of the average estimatefrom the true function’s mean

Expected squared deviation of ourestimate around its mean

2)(and0)(;)( VarEXfYModel:

16

Further Bias Decomposition

• For linear models (eg. Ridge), bias can be further decomposed:

* is the best fitting linear approximationAverage EstimationBias

For standard linear regression, Estimation Bias = 0

2))((minarg XXfE T

Average ModelBias

17

Model Fitting

Graphical representation of bias & variance

Hypothesis SpaceModel Space(basic linear regression)

Regularized Model Space(ridge regression)

Truth

Realization

Model BiasEstimation

Bias

EstimationVariance

Shrunken fit

Closest fit(given our observation)

Closest fitIn population(if epsilon=0)

18

Bias & Variance Decomposition Examples

• kNN Regression

• Linear Regression

Linear weights on y:

Averaging over the training set:

1N

Err(x i)i

2

1N

[ f (x i) Eˆ f (i

x i)]2

pN

2

19

Regression with squared error loss

Classification with 0-1 loss

Prediction error

Bias2

Variance

-- + -- = -- -- + -- = --

-- + -- <> -- -- + -- <> -- Bias-Variance different for 0-1 loss than for squared error loss

Estimation errors on the right side of the boundary don’t hurt!

Simulated Example of Bias Variance Decomposition

20

Optimism of The Training Error Rate

• Typically: training error rate < true error (same data is being used to fit the method

and assess its error)

<

overly optimistic

Err E[L(Y, ˆ f (X))]

err 1N

L(y i, ˆ f (x i))i1

N

21

Estimating Test Error

• Can we estimate the discrepancy between err and Err?

Adjustment for optimism of training error

extra-sample error

Errin --- In-sample error:Expectation over N new

responses at each xi

22

Optimism

Summary: 1

2 ˆ ,N

in y i ii

Err E err Cov y yN

• For linear fit with d indep inputs/basis funcs:

optimism linearly with # d Optimism as training sample size

22in yErr E err d

N

for squared error, 0-1 and other loss functions:

1

2 ˆ ,N

i ii

op Cov y yN

optimism: in yop Err E err

23

Ways to Estimate Prediction Error

• In-sample error estimates: AIC BIC MDL SRM

• Extra-sample error estimates: Cross-Validation

• Leave-one-out• K-fold

Bootstrap

24

• General form of the in-sample estimate:

• For linear fit : 22 ˆ , so called statisticp p

dC err CN

poerrin ˆrrE

Estimates of In-Sample Prediction Error

25

AIC & BIC

Similarly: Akaike Information Criterion (AIC)

AIC 2N

loglik 2dN

Bayesian Information Criterion (BIC)

dNCBI )(logloglik2

26

AIC & BIC

AIC LL(Data | MLE params) (# of parameters)

BIC LL(Data | MLE params) logN

2(# of parameters)

27

MDL(Minimum Description Length)

• Regularity ~ Compressibility• Learning ~ Finding regularities

Learning model

Real model

InputSamples

Rn

PredictionsR1

=?

Real classR1

error

28

MDL(Minimum Description Length)

• Regularity ~ Compressibility• Learning ~ Finding regularities

length logPr(y |,M, X) logPr( | M)

Length of transmitting the discrepancy given the model + optimal coding under the given model

Description of the model under optimal coding

MDL principle: choose the model with the minimum description length

Equivalent to maximizing the posterior:

Pr(y |, M, X)Pr( | M)

29

SRM with VC (Vapnik-Chernovenkis) Dimension

• Vapnik showed that with probability 1-

Errtrue Errtrain 2 1 1

4 Errtrain

As h increases

A method of selecting a class F from a family of nested classes

h VC dimension (measure of f 's power)

where a1

h log(a2N /h) 1 log( /4)N

30

Errin Estimation

• A trade-off between the fit to the data and the model complexity

AIC err 2dN

ˆ e

dNCBI )(logloglik2

MDL length logPr(y |,M, X) logPr( | M)

VC : Errtrue Errtrain 2 1 1

4 Errtrain

31

Estimation of Extra-Sample Err

• Cross Validation

• Bootstrap

32

CV () 1N

L y i, ˆ f (i)(x i,) i1

N

……

test train

K-fold

Cross-Validation

33

How many folds?

Variance decreases bias decreases

Computation increases

k increases

k foldLeave-one-out

34

Cross-Validation: Choosing K

Popular choices for K: 5,10,N

35

Generalized Cross-Validation

• LOOCV can be computational expensive for linear fitting with large N

• GCV provides a computationally cheaper approximation

• Linear fitting

• For linear fitting under squared-error loss:

ˆ y Sy (S is a smoother matrix)

1N

yi ˆ f i(xi) i1

N

2

1N

yi ˆ f (xi)1 Sii

i1

N

2

Sii i'th diagonal element of S

trace(S)N

GCV 1N

yi ˆ f (xi)1

i1

N

2

36

Bootstrap: Main Concept“The bootstrap is a computer-based

method of statistical inference that can answer many real statistical questions without formulas”

(An Introduction to the Bootstrap, Efron and Tibshirani, 1993)

Step 1: Draw samples with replacement

Step 2: Calculate the statistic

37

How is it coming

Sampling distribution of sample mean

x In practice cannot afford

large number of random samples

The theory tells us the sampling distribution

The sample stands for the populationand the distribution of in many resamples stands for the sampling

distribution

x

38

Bootstrap: Error Estimation with Errboot

Vˆ a r[S(Z)] 1

B 1S(Z*b ) S

* b1

B

S*

1B

S(Z*b )b1

N

Var ˆ F [S(Z)] Depends on the unknown true

distribution F

A straightforward application of bootstrap to error prediction

Eˆ r rboot 1B

1N

L(y i, ˆ f *b (x i))i1

N

b1

B

39

Bootstrap: Error Estimation with Err(1)

A CV-inspired improvement on Errboot

Eˆ r r(1) 1N

1C i

L(y i, ˆ f *b (x i))b C i

i1

N

40

Bootstrap: Error Estimation with Err(.632)

An improvement on Err(1) in light-fitting cases

Eˆ r r(.632) .368err .632Eˆ r r(1)

N # of datapoints Z (z1,...,zn )

Probability of zi NOT being chosen when 1 point is uniformly sampled from Z : 1 - 1N

Probability of z i NOT being chosen when Z is sampled N times : 1 - 1N

N

Probability of zi being chosen AT LEAST once when Z is sampled N times: 1 1 - 1N

N

1 e 1 Eˆ r r(.632) err .632(Eˆ r r(1) err)

0.632 .368err .632Eˆ r r(1)

?

41

Bootstrap: Error Estimation with Err(.632+)

An improvement on Err(.632) by adaptively accounting for overfitting

• Depending on the amount of overfitting, the best error estimate is as little as Err(.632) , or as much as Err(1), or something in between

• Err(.632+) is like Err(.632) with adaptive weights, with Err(1) weighted at least .632

• Err(.632+) adaptively mixes training error and leave-one-out error using the relative overfitting rate (R)

42

Bootstrap: Error Estimation with Err(.632+)

Eˆ r r(.632) ranges from Eˆ r r(.632) if there is minimal overfitting (R0),

to Eˆ r r(1) if there is maximal overfitting (R1)

43

Cross Validation & Bootstrap

• Why bother with cross-validation and bootstrap when analytical estimates are known?

1) AIC, BIC, MDL, SRM all requires knowledge of d, which is difficult to attain in most situations.

2) Bootstrap and cross validation gives similar results to above but also applicable in more complex situation.

3) Estimating the noise variance requires a roughly working model, cross validation and bootstrap will work well even if the model is far from correct.

44

Conclusion• Test error plays crucial roles in model selection• AIC, BIC and SRMVC have the advantage that you only need the

training error• If VC-dimension is known, then SRM is a good method for model

selection – requires much less computation than CV and bootstrap, but is wildly conservative

• Methods like CV, Bootstrap give tighter error bounds, but might have more variance

• Asymptotically AIC and Leave-one-out CV should be the same• Asymptotically BIC and a carefully chosen k-fold should be the same • BIC is what you want if you want the best structure instead of the

best predictor• Bootstrap has much wider applicability than just estimating

prediction error