CS229 Problem Set #1 Solutions 1

CS 229, Public Course

Problem Set #1 Solutions: Supervised Learning

1. Newton’s method for computing least squares

In this problem, we will prove that if we use Newton’s method solve the least squares
optimization problem, then we only need one iteration to converge to θ∗ .
Pm
(a) Find the Hessian of the cost function J(θ) = 21 i=1 (θT x(i) − y (i) )2 .
Answer: As shown in the class notes
m
∂J(θ) X T (i) (i)
= (θ x − y (i) )xj .
∂θj i=1

So
m
∂ 2 J(θ) X ∂ (i)
= (θT x(i) − y (i) )xj
∂θj ∂θk i=1
∂θ k
m
(i) (i)
X
= xj xk = (X T X)jk
i=1

Therefore, the Hessian of J(θ) is H = X T X. This can also be derived by simply applying
rules from the lecture notes on Linear Algebra.
(b) Show that the first iteration of Newton’s method gives us θ⋆ = (X T X)−1 X T ~y , the
solution to our least squares problem.
Answer: Given any θ(0) , Newton’s method finds θ(1) according to

θ(1) = θ(0) − H −1 ∇θ J(θ(0) )

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

Therefore, no matter what θ(0) we pick, Newton’s method always finds θ⋆ after one
iteration.

2. Locally-weighted logistic regression

In this problem you will implement a locally-weighted version of logistic regression, where
we weight different training examples differently according to the query point. The locally-
weighted logistic regression problem is to maximize
m
λ X h i
ℓ(θ) = − θT θ + w(i) y (i) log hθ (x(i) ) + (1 − y (i) ) log(1 − hθ (x(i) )) .
2 i=1
CS229 Problem Set #1 Solutions 2

The − λ2 θT θ here is what is known as a regularization parameter, which will be discussed

in a future lecture, but which we include here because it is needed for Newton’s method to
perform well on this task. For the entirety of this problem you can use the value λ = 0.0001.
Using this definition, the gradient of ℓ(θ) is given by

∇θ ℓ(θ) = X T z − λθ

where z ∈ Rm is defined by
zi = w(i) (y (i) − hθ (x(i) ))
and the Hessian is given by
H = X T DX − λI
where D ∈ Rm×m is a diagonal matrix with

Dii = −w(i) hθ (x(i) )(1 − hθ (x(i) ))

For the sake of this problem you can just use the above formulas, but you should try to
derive these results for yourself as well.
Given a query point x, we choose compute the weights

||x − x(i) ||2

 
w(i) = exp − .
2τ 2
Much like the locally weighted linear regression that was discussed in class, this weighting
scheme gives more when the “nearby” points when predicting the class of a new example.

(a) Implement the Newton-Raphson algorithm for optimizing ℓ(θ) for a new query point
x, and use this to predict the class of x.
The q2/ directory contains data and code for this problem. You should implement
the y = lwlr(X train, y train, x, tau) function in the lwlr.m file. This func-
tion takes as input the training set (the X train and y train matrices, in the form
described in the class notes), a new query point x and the weight bandwitdh tau.
Given this input the function should 1) compute weights w(i) for each training exam-
ple, using the formula above, 2) maximize ℓ(θ) using Newton’s method, and finally 3)
output y = 1{hθ (x) > 0.5} as the prediction.
We provide two additional functions that might help. The [X train, y train] =
load data; function will load the matrices from files in the data/ folder. The func-
tion plot lwlr(X train, y train, tau, resolution) will plot the resulting clas-
sifier (assuming you have properly implemented lwlr.m). This function evaluates the
locally weighted logistic regression classifier over a large grid of points and plots the
resulting prediction as blue (predicting y = 0) or red (predicting y = 1). Depending
on how fast your lwlr function is, creating the plot might take some time, so we
recommend debugging your code with resolution = 50; and later increase it to at
least 200 to get a better idea of the decision boundary.
Answer: Our implementation of lwlr.m:
function y = lwlr(X_train, y_train, x, tau)

m = size(X_train,1);
n = size(X_train,2);
CS229 Problem Set #1 Solutions 3

theta = zeros(n,1);

% compute weights
w = exp(-sum((X_train - repmat(x’, m, 1)).^2, 2) / (2*tau));

% perform Newton’s method

g = ones(n,1);
while (norm(g) > 1e-6)
h = 1 ./ (1 + exp(-X_train * theta));
g = X_train’ * (w.*(y_train - h)) - 1e-4*theta;
H = -X_train’ * diag(w.*h.*(1-h)) * X_train - 1e-4*eye(n);
theta = theta - H \ g;
end

% return predicted y
y = double(x’*theta > 0);
(b) Evaluate the system with a variety of different bandwidth parameters τ . In particular,
try τ = 0.01, 0.050.1, 0.51.0, 5.0. How does the classification boundary change when
varying this parameter? Can you predict what the decision boundary of ordinary
(unweighted) logistic regression would look like?
Answer: These are the resulting decision boundaries, for the different values of τ .

tau = 0.01 tau = 0.05 tau = 0.1

tau = 0.5 tau = 0.5 tau = 5

For smaller τ , the classifier appears to overfit the data set, obtaining zero training error,
but outputting a sporadic looking decision boundary. As τ grows, the resulting deci-
sion boundary becomes smoother, eventually converging (in the limit as τ → ∞ to the
unweighted linear regression solution).

3. Multivariate least squares

So far in class, we have only considered cases where our target variable y is a scalar value.
Suppose that instead of trying to predict a single output, we have a training set with
CS229 Problem Set #1 Solutions 4

multiple outputs for each example:

{(x(i) , y (i) ), i = 1, . . . , m}, x(i) ∈ Rn , y (i) ∈ Rp .

Thus for each training example, y (i) is vector-valued, with p entries. We wish to use a linear
model to predict the outputs, as in least squares, by specifying the parameter matrix Θ in

y = ΘT x,

where Θ ∈ Rn×p .

(a) The cost function for this case is

m p 2
1 X X  T (i) (i)
J(Θ) = (Θ x )j − yj .
2 i=1 j=1

Write J(Θ) in matrix-vector notation (i.e., without using any summations). [Hint:
Start with the m × n design matrix

— (x(1) )T —
 
 — (x(2) )T — 
X= ..
 

 . 
— (x(m) )T —

and the m × p target matrix

(y (1) )T
 
— —
 — (y (2) )T — 
Y = ..
 

 . 
— (y (m) )T —

and then work out how to express J(Θ) in terms of these matrices.]
Answer: The objective function can be expressed as
1
tr (XΘ − Y )T (XΘ − Y ) .


J(Θ) =
2
To see this, note that
1
tr (XΘ − Y )T (XΘ − Y )


J(Θ) =
2
1X
XΘ − Y )T (XΘ − Y ) ii


=
2 i
1 XX
= (XΘ − Y )2ij
2 i j
m p
1 X X  T (i) (i)
2
= (Θ x )j − yj
2 i=1 j=1
CS229 Problem Set #1 Solutions 5

(b) Find the closed form solution for Θ which minimizes J(Θ). This is the equivalent to
the normal equations for the multivariate case.
Answer: First we take the gradient of J(Θ) with respect to Θ.
 
1 T


∇Θ J(Θ) = ∇Θ tr (XΘ − Y ) (XΘ − Y )
2
 
1 T T T T T T


= ∇Θ tr Θ X XΘ − Θ X Y − Y XΘ − Y T
2
1
∇Θ tr(ΘT X T XΘ) − tr(ΘT X T Y ) − tr(Y T XΘ) + tr(Y T Y )
 
=
2
1
∇Θ tr(ΘT X T XΘ) − 2tr(Y T XΘ) + tr(Y T Y )
 
=
2
1 T
X XΘ + X T XΘ − 2X T Y

=
2
= X T XΘ − X T Y
Setting this expression to zero we obtain
Θ = (X T X)−1 X T Y.
This looks very similar to the closed form solution in the univariate case, except now Y
is a m × p matrix, so then Θ is also a matrix, of size n × p.
(c) Suppose instead of considering the multivariate vectors y (i) all at once, we instead
(i)
compute each variable yj separately for each j = 1, . . . , p. In this case, we have a p
individual linear models, of the form
(i)
yj = θjT x(i) , j = 1, . . . , p.
(So here, each θj ∈ Rn ). How do the parameters from these p independent least
squares problems compare to the multivariate solution?
Answer: This time, we construct a set of vectors
 (1) 
yj
 (2) 
 yj 
~yj = 
 ..  , j = 1, . . . , p.

 . 
(m)
yj
Then our j-th linear model can be solved by the least squares solution
θj = (X T X)−1 X T ~yj .
If we line up our θj , we see that we have the following equation:
[θ1 θ2 · · · θp ] = (X T X)−1 X T ~y1 (X T X)−1 X T ~y2 · · · (X T X)−1 X T ~yp
 

= (X T X)−1 X T [~y1 ~y2 · · · ~yp ]

= (X T X)−1 X T Y
= Θ.
Thus, our p individual least squares problems give the exact same solution as the multi-
variate least squares.
CS229 Problem Set #1 Solutions 6

4. Naive Bayes
In this problem, we look at maximum likelihood parameter estimation using the naive
Bayes assumption. Here, the input features xj , j = 1, . . . , n to our model are discrete,
binary-valued variables, so xj ∈ {0, 1}. We call x = [x1 x2 · · · xn ]T to be the input vector.
For each training example, our output targets are a single binary-value y ∈ {0, 1}. Our
model is then parameterized by φj|y=0 = p(xj = 1|y = 0), φj|y=1 = p(xj = 1|y = 1), and
φy = p(y = 1). We model the joint distribution of (x, y) according to

p(y) = (φy )y (1 − φy )1−y

Yn
p(x|y = 0) = p(xj |y = 0)
j=1
Yn
= (φj|y=0 )xj (1 − φj|y=0 )1−xj
j=1
Yn
p(x|y = 1) = p(xj |y = 1)
j=1
Yn
= (φj|y=1 )xj (1 − φj|y=1 )1−xj
j=1

Qm
(a) Find the joint likelihood function ℓ(ϕ) = log i=1 p(x(i) , y (i) ; ϕ) in terms of the
model parameters given above. Here, ϕ represents the entire set of parameters
{φy , φj|y=0 , φj|y=1 , j = 1, . . . , n}.
Answer:
m
Y
ℓ(ϕ) = log p(x(i) , y (i) ; ϕ)
i=1
Ym
= log p(x(i) |y (i) ; ϕ)p(y (i) ; ϕ)
i=1
 
m n
(i)
Y Y
= log  p(xj |y (i) ; ϕ) p(y (i) ; ϕ)
i=1 j=1
 
m n
(i)
X X
= log p(y (i) ; ϕ) + log p(xj |y (i) ; ϕ)
i=1 j=1
m
"
X
= y (i) log φy + (1 − y (i) ) log(1 − φy )
i=1

n  
(i) (i)
X
+ xj log φj|y(i) + (1 − xj ) log(1 − φj|y(i) ) 
j=1

(b) Show that the parameters which maximize the likelihood function are the same as
CS229 Problem Set #1 Solutions 7

those given in the lecture notes; i.e., that

Pm (i)
1{x = 1 ∧ y (i) = 0}
φj|y=0 =
i=1
Pm j (i) = 0}
i=1 1{y
Pm (i)
1{x = 1 ∧ y (i) = 1}
φj|y=1 =
i=1
Pm j (i) = 1}
i=1 1{y
Pm
i=1 1{y (i) = 1}
φy = .
m

Answer: The only terms in ℓ(ϕ) which have non-zero gradient with respect to φj|y=0
are those which include φj|y(i) . Therefore,
m  
(i) (i)
X
∇φj|y=0 ℓ(ϕ) = ∇φj|y=0 xj log φj|y(i) + (1 − xj ) log(1 − φj|y(i) )
i=1
m 
(i)
X
= ∇φj|y=0 xj log(φj|y=0 )1{y (i) = 0}
i=1

(i)
+ (1 − xj ) log(1 − φj|y=0 )1{y (i) = 0}
m  
X (i) 1 (i) 1
= xj 1{y (i) = 0} − (1 − xj ) 1{y (i) = 0} .
i=1
φj|y=0 1 − φj|y=0

Setting ∇φj|y=0 ℓ(ϕ) = 0 gives

m  
X (i) 1 (i) (i) 1 (i)
0 = xj 1{y = 0} − (1 − xj ) 1{y = 0}
i=1
φj|y=0 1 − φj|y=0
m  
(i) (i)
X
= xj (1 − φj|y=0 )1{y (i) = 0} − (1 − xj )φj|y=0 1{y (i) = 0}
i=1
m  
(i)
X
= (xj − φj|y=0 )1{y (i) = 0}
i=1
m   m
(i)
X X
= xj · 1{y (i) = 0} − φj|y=0 1{y (i) = 0}
i=1 i=1
m   m
(i)
X X
= 1{xj = 1 ∧ y (i) = 0} − φj|y=0 1{y (i) = 0}.
i=1 i=1

We then arrive at our desired result

Pm (i)
1{x = 1 ∧ y (i) = 0}
φj|y=0 =
i=1
Pm j (i) = 0}
i=1 1{y

The solution for φj|y=1 proceeds in the identical manner.

CS229 Problem Set #1 Solutions 8

To solve for φy ,
m 
X 
∇φy ℓ(ϕ) = ∇φy y (i) log φy + (1 − y (i) ) log(1 − φy )
i=1
m  
X 1 1
= y (i) − (1 − y (i) )
i=1
φy 1 − φy

Then setting ∇φy = 0 gives us

m  
X 1 1
0 = y (i) − (1 − y (i) )
i=1
φy 1 − φy
m 
X 
= y (i) (1 − φy ) − (1 − y (i) )φy
i=1
m
X m
X
= y (i) − φy .
i=1 i=1

Therefore, Pm
1{y (i) = 1}
i=1
φy = .
m
(c) Consider making a prediction on some new data point x using the most likely class
estimate generated by the naive Bayes algorithm. Show that the hypothesis returned
by naive Bayes is a linear classifier—i.e., if p(y = 0|x) and p(y = 1|x) are the class
probabilities returned by naive Bayes, show that there exists some θ ∈ Rn+1 such
that  
1
p(y = 1|x) ≥ p(y = 0|x) if and only if θT ≥ 0.
x
(Assume θ0 is an intercept term.)
Answer:
p(y = 1|x) ≥ p(y = 0|x)
p(y = 1|x)
⇐⇒ ≥1
p(y = 0|x)
Q 
n
j=1 p(x j |y = 1) p(y = 1)
⇐⇒ Q  ≥1
n
j=1 p(xj |y = 0) p(y = 0)
Q 
n xj 1−xj
j=1 (φ j|y=0 ) (1 − φ j|y=0 ) φy
⇐⇒ Q  ≥1
n xj (1 − φ 1−xj (1 − φ )
(φ
j=1 j|y=1 ) j|y=1 ) y
n 
φj|y=1 1 − φj|y=0
     
X φy
⇐⇒ xj log + (1 − xj ) log + log ≥0
j=1
φj|y=0 1 − φj |y = 0 1 − φy
n n
(φj|y=1 )(1 − φj|y=0 ) 1 − φj|y=1
  X    
X φy
⇐⇒ xj log + log + log ≥0
j=1
(φj|y=0 )(1 − φj|y=1 ) j=1
1 − φj|y=0 1 − φy
 
T 1
⇐⇒ θ ≥ 0,
x
CS229 Problem Set #1 Solutions 9

where
n
1 − φj|y=1
   
X φy
θ0 = log + log
j=1
1 − φj|y=0 1 − φy
(φj|y=1 )(1 − φj|y=0 )
 
θj = log , j = 1, . . . , n.
(φj|y=0 )(1 − φj|y=1 )

5. Exponential family and the geometric distribution

(a) Consider the geometric distribution parameterized by φ:

p(y; φ) = (1 − φ)y−1 φ, y = 1, 2, 3, . . . .

Show that the geometric distribution is in the exponential family, and give b(y), η,
T (y), and a(η).
Answer:

p(y; φ) = (1 − φ)y−1 φ
exp log(1 − φ)y−1 + log φ
 
=
= exp [(y − 1) log(1 − φ) + log φ]
  
1−φ
= exp y log(1 − φ) − log
φ

Then

b(y) = 1
η = log(1 − φ)
T (y) = y
   η 
1−φ e
a(η) = log = log ,
φ 1 − eη

where the last line follows becuase η = log(1 − φ) ⇒ eη = 1 − φ ⇒ φ = 1 − eη .

(b) Consider performing regression using a GLM model with a geometric response vari-
able. What is the canonical response function for the family? You may use the fact
that the mean of a geometric distribution is given by 1/φ.
Answer:
1 1
g(η) = E[y; φ] = = .
φ 1 − eη

(c) For a training set {(x(i) , y (i) ); i = 1, . . . , m}, let the log-likelihood of an example
be log p(y (i) |x(i) ; θ). By taking the derivative of the log-likelihood with respect to
θj , derive the stochastic gradient ascent rule for learning using a GLM model with
goemetric responses y and the canonical response function.
Answer: The log-likelihood of an example (x(i) , y (i) ) is defined as ℓ(θ) = log p(y (i) |x(i) ; θ).
To derive the stochastic gradient ascent rule, use the results from previous parts and the
standard GLM assumption that η = θT x.
CS229 Problem Set #1 Solutions 10

" T (i)
!!#
T (i) (i) eθ x
ℓi (θ) = log exp θ x ·y − log
1 − eθT x(i)
   
T (i) (i) 1
= log exp θ x · y − log
e−θT x(i) − 1
 T (i) 
= θT x(i) · y (i) + log e−θ x − 1
T
x(i)
∂ (i) e−θ (i)
ℓi (θ) = xj y (i) + (−xj )
∂θj e−θT x(i) − 1
(i) 1 (i)
= xj y (i) − x
1 − e T x(i) j
−θ
 
(i) 1 (i)
= y − xj .
1 − eθT x(i)
Thus the stochastic gradient ascent update rule should be

∂ℓi (θ)
θj := θj + α ,
∂θj

which is  
1 (i)
θj := θj + α y (i) − xj .
1 − eθT x(i)