CS229 Problem Set #2 Solutions 1

CS 229, Public Course

Problem Set #2 Solutions: Kernels, SVMs, and
Theory

1. Kernel ridge regression

In contrast to ordinary least squares which has a cost function
m
1 X T (i)
J(θ) = (θ x − y (i) )2 ,
2 i=1

we can also add a term that penalizes large weights in θ. In ridge regression, our least
squares cost is regularized by adding a term λkθk2 , where λ > 0 is a fixed (known) constant
(regularization will be discussed at greater length in an upcoming course lecutre). The ridge
regression cost function is then
m
1 X T (i) λ
J(θ) = (θ x − y (i) )2 + kθk2 .
2 i=1 2

(a) Use the vector notation described in class to find a closed-form expreesion for the
value of θ which minimizes the ridge regression cost function.
Answer: Using the design matrix notation, we can rewrite J(θ) as

1 λ
J(θ) = (Xθ − ~y )T (Xθ − ~y ) + θT θ.
2 2
Then the gradient is
∇θ J(θ) = X T Xθ − X T ~y + λθ.
Setting the gradient to 0 gives us

0 = X T Xθ − X T ~y + λθ
θ = (X T X + λI)−1 X T ~y .

(b) Suppose that we want to use kernels to implicitly represent our feature vectors in a
high-dimensional (possibly infinite dimensional) space. Using a feature mapping φ,
the ridge regression cost function becomes
m
1X T λ
J(θ) = (θ φ(x(i) ) − y (i) )2 + kθk2 .
2 i=1 2

Making a prediction on a new input xnew would now be done by computing θT φ(xnew ).
Show how we can use the “kernel trick” to obtain a closed form for the prediction
on the new input without ever explicitly computing φ(xnew ). You may assume that
the parameter vector
Pm θ can be expressed as a linear combination of the input feature
vectors; i.e., θ = i=1 αi φ(x(i) ) for some set of parameters αi .
CS229 Problem Set #2 Solutions 2

[Hint: You may find the following identity useful:

(λI + BA)−1 B = B(λI + AB)−1 .
If you want, you can try to prove this as well, though this is not required for the
problem.]
Answer: Let Φ be the design matrix associated with the feature vectors φ(x(i) ). Then
from parts (a) and (b),

−1 T
θ = ΦT Φ + λI Φ ~y
−1
= ΦT ΦΦT + λI


~y
= ΦT (K + λI)−1 ~y .
where K is the kernel matrix for the training set (since Φi,j = φ(x(i) )T φ(x(j) ) = Kij .)
To predict a new value ynew , we can compute
~ynew = θT φ(xnew )
= ~y T (K + λI)−1 Φφ(xnew )
Xm
= αi K(x(i) , xnew ).
i=1
−1
where α = (K + λI) ~y . All these terms can be efficiently computing using the kernel
function.
To prove the identity from the hint, we left-multiply by λ(I + BA) and right-multiply by
λ(I + AB) on both sides. That is,
(λI + BA)−1 B = B(λI + AB)−1
B = (λI + BA)B(λI + AB)−1
B(λI + AB) = (λI + BA)B
λB + BAB = λB + BAB.
This last line clearly holds, proving the identity.

2. ℓ2 norm soft margin SVMs

In class, we saw that if our data is not linearly separable, then we need to modify our
support vector machine algorithm by introducing an error margin that must be minimized.
Specifically, the formulation we have looked at is known as the ℓ1 norm soft margin SVM.
In this problem we will consider an alternative method, known as the ℓ2 norm soft margin
SVM. This new algorithm is given by the following optimization problem (notice that the
slack penalties are now squared):
Pm
minw,b,ξ 21 kwk2 + C2 i=1 ξi2
.
s.t. y (i) (wT x(i) + b) ≥ 1 − ξi , i = 1, . . . , m
(a) Notice that we have dropped the ξi ≥ 0 constraint in the ℓ2 problem. Show that these
non-negativity constraints can be removed. That is, show that the optimal value of
the objective will be the same whether or not these constraints are present.
Answer: Consider a potential solution to the above problem with some ξ < 0. Then
the constraint y (i) (wT x(i) + b) ≥ 1 − ξi would also be satisfied for ξi = 0, and the
objective function would be lower, proving that this could not be an optimal solution.
CS229 Problem Set #2 Solutions 3

(b) What is the Lagrangian of the ℓ2 soft margin SVM optimization problem?
Answer:
m m
1 T CX 2 X
L(w, b, ξ, α) = w w + ξ − αi [y (i) (wT x(i) + b) − 1 + ξi ],
2 2 i=1 i i=1

where αi ≥ 0 for i = 1, . . . , m.
(c) Minimize the Lagrangian with respect to w, b, and ξ by taking the following gradients:
∇w L, ∂L T
∂b , and ∇ξ L, and then setting them equal to 0. Here, ξ = [ξ1 , ξ2 , . . . , ξm ] .
Answer: Taking the gradient with respect to w, we get
m
X
0 = ∇w L = w − αi y (i) x(i) ,
i=1

which gives us
m
X
w= αi y (i) x(i) .
i=1

Taking the derivative with respect to b, we get

m
∂L X
0= =− αi y (i) ,
∂b i=1

giving us
m
X
0= αi y (i) .
i=1

Finally, taking the gradient with respect to ξ, we have

0 = ∇ξ L = Cξ − α,

where α = [α1 , α2 , . . . , αm ]T . Thus, for each i = 1, . . . , m, we get

0 = Cξi − αi ⇒ Cξi = αi .

(d) What is the dual of the ℓ2 soft margin SVM optimization problem?
CS229 Problem Set #2 Solutions 4

Answer: The objective function for the dual is

W (α) = min L(w, b, ξ, α)

w,b,ξ
m Xm m
1 X 1 X αi 2
= (αi y (i) x(i) )T (αj y (j) x(j) ) + ξ
2 i=1 j=1
2 i=1 ξi i
  T 
m
X Xm
− αi y (i)  αj y (j) x(j)  x(i) + b − 1 + ξi 
   
i=1 j=1

m m m
1 XX 1X
= − αi αj y (i) y (j) (x(i) )T x(j) + αi ξi
2 i=1 j=1 2 i=1
m
! m m
X X X
(i)
− αi y b+ αi − αi ξi
i=1 i=1 i=1
m m m m
X 1 XX 1X
= αi − αi αj y (i) y (j) (x(i) )T x(j) − αi ξi
i=1
2 i=1 j=1
2 i=1
m m X m m
X 1 X 1 X αi2
= αi − αi αj y (i) y (j) (x(i) )T x(j) − .
i=1
2 i=1 j=1
2 i=1 C

Then the dual formulation of our problem is

Pm Pm Pm Pm α2i
maxα i=1 αi − 12 i=1 j=1 αi αj y (i) y (j) (x(i) )T x(j) − 1
2 i=1 C
s.t. αi ≥ 0, i = 1, . . . , m .
P m (i)
i=1 αi y =0

3. SVM with Gaussian kernel

Consider the task of training a support vector machine using the Gaussian kernel K(x, z) =
exp(−kx − zk2 /τ 2 ). We will show that as long as there are no two identical points in the
training set, we can always find a value for the bandwidth parameter τ such that the SVM
achieves zero training error.

(a) Recall from class that the decision function learned by the support vector machine
can be written as
m
X
f (x) = αi y (i) K(x(i) , x) + b.
i=1

Assume that the training data {(x , y (1) ), . . . , (x(m) , y (m) )} consists of points which
(1)

are separated by at least a distance of ǫ; that is, ||x(j) − x(i) || ≥ ǫ for any i 6= j.
Find values for the set of parameters {α1 , . . . , αm , b} and Gaussian kernel width τ
such that x(i) is correctly classified, for all i = 1, . . . , m. [Hint: Let αi = 1 for all i
and b = 0. Now notice that for y ∈ {−1, +1} the prediction on x(i) will be correct if
|f (x(i) ) − y (i) | < 1, so find a value of τ that satisfies this inequality for all i.]
CS229 Problem Set #2 Solutions 5

Answer: First we set αi = 1 for all i = 1, . . . , m and b = 0. Then, for a training

example {x(i) , y (i) }, we get
 
  X m 
f (x(i) ) − y (i)  =  y (j) K(x(j) , x(i) ) − y (i) 
   
 j=1 
 
 m (j)
X   
y exp −kx(j) − x(i) k2 /τ 2 − y (i) 

= 
 j=1 
 
   
 (i) X (j)
y exp kx(j) − x(i) k2 /τ 2 − y (i) 

= y +
 j6=i 
 
X  
y (j) exp −kx(j) − x(i) k2 /τ 2 

= 
 j6=i 
X   
≤ y (j) exp −kx(j) − x(i) k2 /τ 2 

j6=i
X    
= y (j)  · exp kx(j) − x(i) k2 /τ 2

j6=i
X  
= exp −kx(j) − x(i) k2 /τ 2
j6=i
X
exp −ǫ2 /τ 2


≤
j6=i

(m − 1) exp −ǫ2 /τ 2 .


=

The first inequality comes from repeated application of the triangle inequality |a + b| ≤
|a| + |b|, and the second inequality (1) from the assumption that ||x(j) − x(i) || ≥ ǫ for all
i 6= j. Thus we need to choose a γ such that

(m − 1) exp(−ǫ2 /τ 2 ) < 1,

or
ǫ
τ< .
log(m − 1)
By choosing, for example, τ = ǫ/ log m we are done.
(b) Suppose we run a SVM with slack variables using the parameter τ you found in part
(a). Will the resulting classifier necessarily obtain zero training error? Why or why
not? A short explanation (without proof) will suffice.
Answer: The classifier will obtain zero training error. The SVM without slack variables
will always return zero training error if it is able to find a solution, so all that remains to
be shown is that there exists at least one feasible point.
Consider the constraint y (i) (wT x(i) + b) for some i, and let b = 0. Then

y (i) (wT x(i) + b) = y (i) · f (x(i) ) > 0

since f (x(i) ) and y (i) have the same sign, and shown above. Therefore, as we choose all
the αi ’s large enough, y (i) (wT x(i) + b) > 1, so the optimization problem is feasible.
CS229 Problem Set #2 Solutions 6

(c) Suppose we run the SMO algorithm to train an SVM with slack variables, under
the conditions stated above, using the value of τ you picked in the previous part,
and using some arbitrary value of C (which you do not know beforehand). Will this
necessarily result in a classifier that achieve zero training error? Why or why not?
Again, a short explanation is sufficient.
Answer: The resulting classifier will not necessarily
Pm obtain zero training error. The C
parameter controls the relative weights of the (C i=1 ξi ) and ( 12 ||w||2 ) terms of the SVM
training objective. If the C parameter is sufficiently small, then the former component will
have relatively little contribution to the objective. In this case, a weight vector which has
a very small norm but does not achieve zero training error may achieve a lower objective
value than one which achieves zero training error. For example, you can consider the
extreme case where C = 0, and the objective is just the norm of w. In this case, w = 0 is
the solution to the optimization problem regardless of the choise of τ , this this may not
obtain zero training error.
4. Naive Bayes and SVMs for Spam Classification
In this question you’ll look into the Naive Bayes and Support Vector Machine algorithms
for a spam classification problem. However, instead of implementing the algorithms your-
self, you’ll use a freely available machine learning library. There are many such libraries
available, with different strengths and weaknesses, but for this problem you’ll use the
WEKA machine learning package, available at http://www.cs.waikato.ac.nz/ml/weka/.
WEKA implements many standard machine learning algorithms, is written in Java, and
has both a GUI and a command line interface. It is not the best library for very large-scale
data sets, but it is very nice for playing around with many different algorithms on medium
size problems.
You can download and install WEKA by following the instructions given on the website
above. To use it from the command line, you first need to install a java runtime environ-
ment, then add the weka.jar file to your CLASSPATH environment variable. Finally, you
can call WEKA using the command:
java <classifier> -t <training file> -T <test file>
For example, to run the Naive Bayes classifier (using the multinomial event model) on our
provided spam data set by running the command:
java weka.classifiers.bayes.NaiveBayesMultinomial -t spam train 1000.arff -T spam test.arff

The spam classification dataset in the q4/ directory was provided courtesy of Christian
Shelton ([email protected]). Each example corresponds to a particular email, and each
feature correspondes to a particular word. For privacy reasons we have removed the actual
words themselves from the data set, and instead label the features generically as f1, f2, etc.
However, the data set is from a real spam classification task, so the results demonstrate the
performance of these algorithms on a real-world problem. The q4/ directory actually con-
tains several different training files, named spam train 50.arff, spam train 100.arff,
etc (the “.arff” format is the default format by WEKA), each containing the corresponding
number of training examples. There is also a single test set spam test.arff, which is a
hold out set used for evaluating the classifier’s performance.
(a) Run the weka.classifiers.bayes.NaiveBayesMultinomial classifier on the dataset
and report the resulting error rates. Evaluate the performance of the classifier using
each of the different training files (but each time using the same test file, spam test.arff).
Plot the error rate of the classifier versus the number of training examples.
CS229 Problem Set #2 Solutions 7

(b) Repeat the previous part, but using the weka.classifiers.functions.SMO classifier,
which implements the SMO algorithm to train an SVM. How does the performance
of the SVM compare to that of Naive Bayes?
Answer: Using the above command line arguments to run the classifier, we obtain
the following error rates for the two algorithms:

7 Support Vector Machine

Naive Bayes
6

5
Error (%)

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Num Training Examples

For small amounts of data, Naive Bayes performs better than the Support Vector Machine.
However, as the amount of data grows, the SVM achieves a better error rate.

5. Uniform convergence
In class we proved that for any finite set of hypotheses H = {h1 , . . . , hk }, if we pick the
hypothesis ĥ that minimizes the training error on a set of m examples, then with probability
at least (1 − δ), r
  1 2k
ε(ĥ) ≤ min ε(hi ) + 2 log ,
i 2m δ
where ε(hi ) is the generalization error of hypothesis hi . Now consider a special case (often
called the realizable case) where we know, a priori, that there is some hypothesis in our
class H that achieves zero error on the distribution from which the data is drawn. Then
we could obviously just use the above bound with mini ε(hi ) = 0; however, we can prove a
better bound than this.

(a) Consider a learning algorithm which, after looking at m training examples, chooses
some hypothesis ĥ ∈ H that makes zero mistakes on this training data. (By our
assumption, there is at least one such hypothesis, possibly more.) Show that with
probability 1 − δ
1 k
ε(ĥ) ≤ log .
m δ
Notice that since we do not have a square root here, this bound is much tighter. [Hint:
Consider the probability that a hypothesis with generalization error greater than γ
makes no mistakes on the training data. Instead of the Hoeffding bound, you might
also find the following inequality useful: (1 − γ)m ≤ e−γm .]
Answer: Let h ∈ H be a hypothesis with true error greater than γ. Then

P (“h predicts correctly”) ≤ 1 − γ,

CS229 Problem Set #2 Solutions 8

so
P (“h predicts correctly m times”) ≤ (1 − γ)m ≤ e−γm .
Applying the union bound,

P (∃h ∈ H, s.t. ε(h) > γ and “h predicts correct m times”) ≤ ke−γm .

We want to make this probability equal to δ, so we set

ke−γm = δ,

which gives us
1 k
γ= log .
m δ
This impiles that with probability 1 − δ,
1 k
ε(ĥ) ≤ log .
m δ

(b) Rewrite the above bound as a sample complexity bound, i.e., in the form: for fixed
δ and γ, for ε(ĥ) ≤ γ to hold with probability at least (1 − δ), it suffices that m ≥
f (k, γ, δ) (i.e., f (·) is some function of k, γ, and δ).
Answer: From part (a), if we take the equation,

ke−γm = δ

and solve for m, we obtain

1 k
m= log .
γ δ
Therefore, for m larger than this, ε(ĥ) ≤ γ will hold with probability at least 1 − δ.