Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: THEend8_
COMP9417 - Machine Learning
Homework 2: Kernel Features & Model Combinations
Introduction In this homework we first take a closer look at feature maps induced by kernels. We then ex-
plore a creative use of the gradient descent method introduced in homework 1. We will show that gradient
descent techniques can be used to construct combinations of models from a base set of models such that the
combination can outperform any single base model.
Points Allocation There are a total of 28 marks.
• Question 1 a): 3 marks
• Question 1 b): 3 marks
• Question 1 c): 3 marks
• Question 1 d): 5 marks
• Question 2 a): 2 marks
• Question 2 b): 2 marks
• Question 2 c): 4 marks
• Question 2 d): 2 marks
• Question 2 e): 2 marks
• Question 2 f): 1 mark
• Question 2 g): 1 mark
What to Submit
• A single PDF file which contains solutions to each question. For each question, provide your solution
in the form of text and requested plots. For some questions you will be requested to provide screen
shots of code used to generate your answer — only include these when they are explicitly asked for.
• .py file(s) containing all code you used for the project, which should be provided in a separate .zip
file. This code must match the code provided in the report.
• You may be deducted points for not following these instructions.
• You may be deducted points for poorly presented/formatted work. Please be neat and make your
solutions clear. Start each question on a new page if necessary.
1
• You cannot submit a Jupyter notebook; this will receive a mark of zero. This does not stop you from
developing your code in a notebook and then copying it into a .py file though, or using a tool such as
nbconvert or similar.
• We will set up a Moodle forum for questions about this homework. Please read the existing questions
before posting new questions. Please do some basic research online before posting questions. Please
only post clarification questions. Any questions deemed to be fishing for answers will be ignored
and/or deleted.
• Please check Moodle announcements for updates to this spec. It is your responsibility to check for
announcements about the spec.
• Please complete your homework on your own, do not discuss your solution with other people in the
course. General discussion of the problems is fine, but you must write out your own solution and
acknowledge if you discussed any of the problems in your submission (including their name(s) and
zID).
• As usual, we monitor all online forums such as Chegg, StackExchange, etc. Posting homework ques-
tions on these site is equivalent to plagiarism and will result in a case of academic misconduct.
• You may not use SymPy or any other symbolic programming toolkits to answer the derivation ques-
tions. This will result in an automatic grade of zero for the relevant question. You must do the
derivations manually.
When and Where to Submit
• Due date: Week 7, Monday March 27th, 2023 by 5pm. Please note that the forum will not be actively
monitored on weekends.
• Late submissions will incur a penalty of 5% per day from the maximum achievable grade. For ex-
ample, if you achieve a grade of 80/100 but you submitted 3 days late, then your final grade will be
80− 3× 5 = 65. Submissions that are more than 5 days late will receive a mark of zero.
• Submission must be made on Moodle, no exceptions.
Page 2
Question 1. Kernel Power
Consider the following 2-dimensional data-set, where y denotes the class of each point.
index x1 x2 y
1 1 0 -1
2 0 1 -1
3 0 -1 -1
4 -1 0 +1
5 0 2 +1
6 0 -2 +1
7 -2 0 +1
Throughout this question, you may use any desired packages to answer the questions.
(a) Use the transformation x = (x1, x2) 7→ (φ1(x), φ2(x)) where φ1(x) = 2x22 − 4x1 + 1 and φ2(x) =
x21 − 2x2 − 3. What is the equation of the best separating hyper-plane in the new feature space?
Provide a plot with the data set and hyperplane clearly shown.
What to submit: a single plot, the equation of the separating hyperplane, a screen shot of your code, a copy
of your code in your .py file for this question.
(b) Fit a hard margin linear SVM to the transformed data-set in the previous part1. What are the
estimated values of (α1, . . . , α7). Based on this, which points are the support vectors? What error
does your computed SVM achieve?
What to submit: the indices of your identified support vectors, the train error of your SVM, the computed
α’s (rounded to 3 d.p.), a screen shot of your code, a copy of your code in your .py file for this question.
(c) Consider now the kernel k(x, z) = (2+x>z)2. Run a hard-margin kernel SVM on the original (un-
transformed) data given in the table at the start of the question. What are the estimated values of
(α1, . . . , α7). Based on this, which points are the support vectors? What error does your computed
SVM achieve?
What to submit: the indices of your identified support vectors, the train error of your SVM, the computed
α’s (rounded to 3 d.p.), a screen shot of your code, a copy of your code in your .py file for this question.
(d) Provide a detailed argument explaining your results in parts (i), (ii) and (iii). Your argument
should explain the similarities and differences in the answers found. In particular, is your answer
in (iii) worse than in (ii)? Why? To get full marks, be as detailed as possible, and use mathematical
arguments or extra plots if necessary.
What to submit: some commentary and/or plots. If you use any code here, provide a screen shot of your code,
and a copy of your code in your .py file for this question.
Question 2. Gradient Descent for Learning Combinations of Models
In this question, we discuss and implement a gradient descent based algorithm for learning combina-
tions of models, which are generally termed ’ensemble models’. The gradient descent idea is a very
powerful one that has been used in a large number of creative ways in machine learning beyond direct
minimization of loss functions as in the previous question.
The Gradient-Combination (GC) algorithm can be described as follows: Let F be a set of base learning
algorithms2. The idea is to combine the base learners in F in an optimal way to end up with a good
1If you are using the SVC class in sklearn, to get a hard-margin svm, you need to set the hyper parameter C to be very large.
2For example, you could take F to be the set of all regression models with a single feature, or alternatively the set of all regression
models with 4 features, or the set of neural networks with 2 layers etc.
Page 3
learning algorithm. Let `(y, yˆ) be a loss function, where y is the target, and yˆ is the predicted value.3
Suppose we have data (xi, yi) for i = 1, . . . , n, which we collect into a single data set D0. We then set
the number of desired base learners to T and proceed as follows:
(I) Initialize f0(x) = 0 (i.e. f0 is the zero function.)
(II) For t = 1, 2, . . . , T :
(GC1) Compute:
rt,i = − ∂
∂f(xi)
n∑
j=1
`(yj , f(xj))
∣∣∣∣
f(xj)=ft−1(xj), j=1,...,n
for i = 1, . . . , n. We refer to rt,i as the i-th pseudo-residual at iteration t.
(GC2) Construct a new pseudo data set, Dt, consisting of pairs: (xi, rt,i) for i = 1, . . . , n.
(GC3) Fit a model to Dt using our base class F . That is, we solve
ht = argmin
f∈F
n∑
i=1
`(rt,i, f(xi))
(GC4) Choose a step-size. This can be done by either of the following methods:
(SS1) Pick a fixed step-size αt = α
(SS2) Pick a step-size adaptively according to
αt = argmin
α
n∑
i=1
`(yi, ft−1(xi) + αht(xi)).
(GC5) Take the step
ft(x) = ft−1(x) + αtht(x).
(III) return fT .
We can view this algorithm as performing (functional) gradient descent on the base class F . Note that
in (GC1), the notation means that after taking the derivative with respect to f(xi), set all occurences
of f(xj) in the resulting expression with the prediction of the current model ft−1(xj), for all j. For
example:
∂
∂x
log(x+ 1)
∣∣∣∣
x=23
=
1
x+ 1
∣∣∣∣
x=23
=
1
24
.
(a) Consider the regression setting where we allow the y-values in our data set to be real numbers.
Suppose that we use squared error loss `(y, yˆ) = 12 (y− yˆ)2. For round t of the algorithm, show that
rt,i = yi − ft−1(xi). Then, write down an expression for the optimization problem in step (GC3)
that is specific to this setting (you don’t need to actually solve it).
What to submit: your working out, either typed or handwritten.
(b) Using the same setting as in the previous part, derive the step-size expression according to the
adaptive approach (SS2).
What to submit: your working out, either typed or handwritten.
3Note that this set-up is general enough to include both regression and classification algorithms.
Page 4
(c) We will now implement the gradient-combination algorithm on a toy dataset from scratch, and we
will use the class of decision stumps (depth 1 decision trees) as our base class (F), and squared error
loss as in the previous parts.4. The following code generates the data and demonstrates plotting
the predictions of a fitted decision tree (more details in q2.py):5
1 np.random.seed(123)
2 X, y = f_sampler(f, 160, sigma=0.2)
3 X = X.reshape(-1,1)
4
5 fig = plt.figure(figsize=(7,7))
6 dt = DecisionTreeRegressor(max_depth=2).fit(X,y) # example model
7 xx = np.linspace(0,1,1000)
8 plt.plot(xx, f(xx), alpha=0.5, color=’red’, label=’truth’)
9 plt.scatter(X,y, marker=’x’, color=’blue’, label=’observed’)
10 plt.plot(xx, dt.predict(xx.reshape(-1,1)), color=’green’, label=’dt’) # plotting
example model
11 plt.legend()
12 plt.show()
13
The figure generated is
Your task is to generate a 5 x 2 figure of subplots showing the predictions of your fitted gradient-
combination model. There are 10 subplots in total, the first should show the model with 5 base
learners, the second subplot should show it with 10 base learners, etc. The last subplot should be
the gradient-combination model with 50 base learners. Each subplot should include the scatter of
4In your implementation, you may make use of sklearn.tree.DecisionTreeRegressor, but all other code must be your
own. You may use NumPy and matplotlib, but do not use an existing implementation of the algorithm if you happen to find one.
5Although we will not cover decision trees until week 4, we are treating the decision tree as a black box algorithm that can be called
using the sklearn implementation. For more on using sklearn models, see Lab 1.
Page 5
data, as well as a plot of the true model (basically, the same as the plot provided above but with
your fitted model in place of dt). Comment on your results, what happens as the number of base
learners is increased? You should do this two times (two 5x2 plots), once with the adaptive step
size, and the other with the step-size taken to be α = 0.1 fixed throughout. There is no need to
split into train and test data here. Comment on the differences between your fixed and adaptive
step-size implementations. How does your model perform on the different x-ranges of the data?
What to submit: two 5 x 2 plots, one for adaptive and one for fixed step size, some commentary, and a screen
shot of your code and a copy of your code in your .py file.
(d) Repeat the analysis in the previous question but with depth 2 decision trees as base learners in-
stead. Provide the same plots. What do you notice for the adaptive case? What about the non-
adaptive case? What to submit: two 5 x 2 plots, one for adaptive and one for fixed step size, some commen-
tary, and a copy of your code in your .py file.
(e) Now, consider the classification setting where y is taken to be an element of {−1, 1}. We consider
the following classification loss: `(y, yˆ) = log(1 + e−yyˆ). For round t of the algorithm, what is the
expression for rt,i? Write down an expression for the optimization problem in step (GC3) that is
specific to this setting (you don’t need to actually solve it).
What to submit: your working out, either typed or handwritten.
(f) Using the same setting as in the previous part, write down an expression for αt using the adaptive
approach in (SS2). Can you solve for αt in closed form? Explain.
What to submit: your working out, either typed or handwritten, and some commentary.
(g) In practice, if you cannot solve for αt exactly, explain how you might implement the algorithm.
Assume that using a constant step-size is not a valid alternative. Be as specific as possible in your
answer. What, if any, are the additional computational costs of your approach relative to using a
constant step size ?
What to submit: some commentary.