Lecutre 10: Generalization

Model generalization
Test error
Bias, variance and complexity
In-sample error
“Bet on sparsity”
Training data
Testing data
Testing error rate
Training error rate
Good performance on testing data, which is
independent from the training data, is most
important for a model.
It serves as the basis in model selection.
Test error
To evaluate prediction accuracy, loss functions are
needed. Continuous Y:
In classification, categorical G (class label):
Except in rare cases (e.g. 1 nearest neighbor), the trained
classifier always gives a probabilistic outcome.
Test error
The log-likelihood can be used as a loss-function for
general response densities, such as the Poisson,
gamma, exponential, log-normal and others.
If Prθ(X)(Y) is the density of Y , indexed by a parameter
θ(X) that depends on the predictor X, then
The 2 makes the log-likelihood loss for the
Gaussian distribution match squared error loss.
Test error
Test error: The expected loss over an
INDEPENDENT test set. The expectation is taken
with regard to everything that’s random - both the
training set and the test set.
In practice it is more feasible to estimate the
testing error given a training set:
Training error is the average
Loss over just the training set:
Test error
Test error for categorical outcome:
Training error:
Goals in model building
(1) Model selection:
Estimating the performance of different
models; choose the best one
(2) Model assessment:
Estimate the prediction error of the chosen
model on new data.
Goals in model building
Ideally, we’d like to have enough data to be
divided into three sets:
Training set: to fit the models
Validation set: to estimate prediction error of
models, for the purpose of model selection
Test set: to assess the generalization error of
the final model
A typical split:
Goals in model building
What’s the difference between the validation set
and the test set?
The validation set is used repeatedly on all models.
The model selection can chase the randomness in
this set. Our selection of the model is based on this
set. In a sense, there is over-fitting in terms of this
set, and the error rate is under-estimated.
The test set should be protected and used only
once to obtain an unbiased error rate.
Goals in model building
In reality, there’s not enough data. How do people
deal with the issue?
Eliminate validation set.
Draw validation set from training set.
Try to achieve generalization error and model
selection. (AIC, BIC, cross-validation ……)
Sometimes, even omit the test set and final
estimation of prediction error; publish the result
and leave testing to later studies.
Bias-variance trade-off
In the continuous outcome case, assume
The expected prediction error in regression is:
EPE ( x0 ) = E[(Y - fˆ ( x0 )) 2 ]
= E[(e 2 + 2e ( f ( x0 ) - fˆ ( x0 )) + ( f ( x0 ) - fˆ ( x0 )) 2 )]
= s 2 + E[( f ( x ) - fˆ ( x )) 2 ]
= s 2 + E[ fˆ ( x0 ) - E ( fˆ ( x0 ))]2 + [ E ( fˆ ( x0 )) - f ( x0 )]2
= s 2 + Var ( fˆ ( x )) + Bias 2 ( fˆ ( x ))
Bias-variance trade-off
Bias-variance trade-off
K-nearest neighbor classifier:
The higher the k, the lower the model complexity
(estimation becomes more global, space
partitioned into larger patches)
Increase k, the variance term decreases, and the
bias term increases. (Here x’s are assumed to be
fixed; randomness only in y)
Bias-variance trade-off
For linear model with p coefficients,
Although h(x0) is dependent on x0, its average
over sample values is p/N
Model complexity is directly associated with p.
Bias-variance trade-off
Bias-variance trade-off
An example. 50 observations, 20 predictors,
uniformly distributed in the hypercube [0, 1]20
Y is 0 if X1 ≤ 1/2 and 1 if X1 > 1/2, and apply k-nearest neighbors.
Red: prediction
Green: squared
Blue: variance
Bias-variance trade-off
An example. 50 observations, 20 predictors,
uniformly distributed in the hypercube [0, 1]20
Red: prediction
Green: squared
Blue: variance
In-sample error
With limited data, we need to approach testing
error (hidden) as much as we can, and/or
perform model selection.
Two general approaches:
(1) Analytical
AIC, BIC, Cp, ….
(2) Resampling-based
Cross-validation, bootstrap, jackknife, ……
In-sample error
Training error:
This is an under-estimate of the true error
Because the same data is used to fit the model
and assess the error.
Err is extra-sample error, because test data
points can come from outside the training set.
In-sample error
We will only know Err when we know the
population. However, we know only the training
sample which is drawn from the population, but
not the population itself.
In-sample error: obtained when new responses
are observed at the same x’s as the training set
In-sample error
New sample
at same x’s
Define the optimism as
For squared error, 0-1, and other loss function, it
can be shown generally that
In-sample error
 The amount by which Err underestimates the
true error depends on how strongly yi affects
its own prediction.
 Generally, Errin is a better estimate of the test
error than the training error, because the insample is a better approximation to the
 So the goal is to estimate the optimism and
add it to the training error. (Cp, AIC, BIC work
this way for models linear in their parameters.)
In-sample error
Cp statistic:
 : tuning parameter
d( ) : effective number of parameters
minimize over 
In-sample error
In-sample error
BIC tends to choose overly simple models when
sample size is small; it chooses the correct model
when sample size approaches infinity.
AIC tends to choose overly complex models when
sample size is large.
The goal is to directly estimate the extra-sample
error (error on an independent test set)
K-fold cross-validation:
Split data into K roughly equal-sized parts
For each of the K parts, fit the model with the
other K-1 parts, and calculate the prediction
error on the part that is left out.
The CV estimate of the prediction error is from
the combination of the K estimates
α is the tuning parameter (different models,
model parameters)
that minimizes CV(α)
Finally, fit all data on the model
CV could
prediction error,
depending on
the learning
Leave-one-out cross-validation (K=N) is
approximately unbiased, yet it has high
K=5 or 10, CV has low variance but more bias.
If the learning curve has large slope at the
training set size, a 5-fold or 10-fold CV can
overestimate the prediction error substantially.
In multi-stage modeling.
5000 genes, 50 normal/50 disease subjects
Select 100 genes
that best correlate
with disease status
Build a multivariate
predictor using the
100 genes.
CV error rate: 3%.
5-fold CV.
The 100 genes don’t
change; model
parameters are tuned
5-fold CV
Split the data.
For each fold –
Select 100 genes using 4/5 of the samples.
Build a classification model using the 4/5
Predict the other 1/5 samples.
Resample the training set to generate B
bootstrap samples.
Fit the model on each of the B samples. Examine
prediction accuracy for each observation from
those models not built from the observation.
For example the leave-one-out bootstrap
estimate of prediction error is
Reminder: in bootstrap, each sample has the
probability of 0.632 of being selected into a resample.
Average number of distinct observations in a
bootstrap sample is 0.632N.
The (upward) bias will behave like a two-fold
The 0.632 estimator can alleviate the bias
Intuitively it pulls the upward biased leave-oneout error rate towards the downward biased
training set error rate.
“Bet on sparsity” principle
L1 penalty yields sparse models.
L2 penalty yields dense models, and is computationally
When underlying model is sparse, L1 penalty does well.
When underlying model is dense, and the training data
size is not extremely large, neither does well because of
the curse of dimensionality.
“Use a procedure that does well in sparse problems,
since no procedure does well in dense problems”
“Bet on sparsity” principle
300 predictors,
50 observations,
Dense scenario:
all 300 betas are non-zero
Sparse scenario:
only 10 betas are non-zero
30 betas are non-zero
“Bet on sparsity” principle
Data generated from dense model –
neither do well.
“Bet on sparsity” principle
Date generated from sparse model – Lasso
does well.
“Bet on sparsity” principle
Date generated from relatively sparse model –
Lasso does better than ridge.

similar documents