DS5220: Supervised Machine Learning and Learning Theory Fall 2022
Problem Set 1
Instructor: Hongyang Ryan Zhang Due: September 30, 2022, 11:59pm
Instructions:
• You are expected to write up the solution on your own. Discussions and collaborations are
encouraged; remember to mention any fellow students you discussed with when you turn in
the solution.
• There are up to three late days for all the problem sets and project submissions. Use them
wisely. After that, the grade depreciates by 20% for every extra day. Late submissions are
considered case by case. Please reach out to the instructor if you cannot meet the deadline.
• Submit your written solutions to Gradescope and upload your code to Canvas. You are
recommended to write up the solution in LaTeX.
Problem 1 (20 points)
(a) (1 point) Calculate V ar(X) when X represents the outcome when a fair coin flip (i.e., E[X] =
1/2).
(b) (1 point) Find the expected value of the sum obtained when n fair coin flips are rolled
independently.
(c) (2 point) For three events A, B, and C, we know that: A and C are independent, B and C are
independent, A and B are disjoint, P(A ∪ C) = 2/3, P(B ∪ C) = 3/4, P(A ∪ B ∪ C) = 11/12.
Find P(A), P(B) and P(C).
(d) (2 points) Consider a test to detect a disease (e.g., COVID-19), assuming that 0.6% of the
population has it. The test is 97% effective in detecting an infected person. However, the
test gives a false-positive result in 1% of cases (meaning that it shows a positive result if the
person is not infected). What is the probability that a person gets a negative test result?
(e) (2 points) If a person tests positive for the disease, what is the probability that they actually
have COVID?
(f) (2 points) If a person tests negative for the disease, what is the probability that they are
infected with COVID?
Along with the tests, data regarding the number of symptoms shown by the patients was also
recorded and is given below. The data was collected from 2 different sources.
No. of Symptoms Patients
1 20
2 20
3 20
4 20
No. of Symptoms Patients
1 70
2 15
3 10
4 5
(g) (2 points) Suppose you pick one patient from each of the above 2 sources independently.
What would be the expected number of symptoms detected in each of them?
(h) (2 points) Prove that V ar(X) = E[X]
2−(E[X])2
. Explain the interpretation of this derivation.
(i) (2 points) Let Y1 and Y2 denote the number of symptoms detected in each of the above two
patients respectively, where Y1, Y2 ∈ [1, 2, 3, 4]. Then calculate the following probabilities: (i)
E[Y1Y2]; (ii) V ar[Y1 − Y2].
(j) (2 points) Among a population of n people, let X be the number of people that test positive.
What is the expectation of X, E[X]? What is the variance of X, V ar[X]? Make sure to
include all the steps in the calculation.
(k) (2 points) Define bias error and variance error. What do you understand by Bias-Variance
trade-off?
Problem 2 (20 points)
(a) (2 points) Show that for any arbitrary matrix X ∈ Rm×n, the matrix XX⊤ is always positive
semi-definite.
(b) Recall that the SVD of a rank-r matrix M has the form
M =
Xr
i=1
σiuiv
T
i
,
where {ui}
r
i=1 denote the left singular vectors, {vi}
r
i=1 denote the right singular vectors, and
{σi}
r
i=1 denote the singular values.
i) (2 points) Let
A =
1 1
1 1
1 −1
.
Calculate the left and right singular vectors {ui}
r
i=1 and {vi}
r
i=1 of A. Then show that
{ui}
r
i=1 and {vi}
r
i=1 are the eigenvectors of AA⊤ and A⊤A.
ii) (5 points) Let M ∈ R
m×n be an arbitrary real-valued rank-r matrix, show that the
eigenvectors of MMT and MTM are {ui}
r
i=1 and {vi}
r
i=1 respectively.
(c) Recall that the best rank-k approximation of M in Frobenius norm is attained by
B =
X
k
i=1
σiuiv
T
i
.
i) (2 points) For the matrix A defined above, calculate the best rank-1 approximation of
A in Frobenius norm. Then find out the approximation error ||M − B||F .
ii) (5 points) Let M ∈ R
m×n be an arbitrary real-valued rank-r matrix. Show that
||M − B||F =
vuut
Xr
i=k+1
σ
2
i
.
(d) (4 points) Write a Python file to verify your calculation in (b-i) and (c-i). You may find the
library numpy.linalg.svd and numpy.linalg.eig useful.
Problem 3 (15 points)
(a) (6 points) For vectors x ∈ R
n
, a ∈ R
n and matrices X ∈ R
n×n
, A ∈ R
n×n
, show the following:
(i) ∂a
Tx
∂x
= a.
(ii) ∂x
T Ax
∂x
= (A + AT )x.
(iii) ∂||y − Ax||2
2
∂x
= 2AT
(Ax − y).
(b) (4 points) You are given a training set {(x1, y1), . . . ,(xn, yn)}, where xi ∈ R
d and yi ∈ R.
Consider the regression problem
min
θ∈Rd
1
n
Xn
i=1
(yi − θ
Txi))2
.
What is the minimizer of the above regression problem? Provide all steps of your derivation.
Feel free to assume that the rank of {xi}
n
i=1 is equal to d.]
(c) (5 points) Let the cost function to minimize is:
J(w) = Xn
i=1
(yi − θ
Txi))2 + λ
X
d
j=0
θj
2
Prove that the vector w
⋆
that minimizes J(w) is:
w
⋆ = (X⊤X + λI)
−1X⊤y,
where X is the n by d design matrix, whose i-th row is xi
, and y = (y1, …, yn)
⊤.
Problem 4 (15 points)
We consider a regression problem for predicting the demand of bike-sharing services in Washington
D.C.1 The prediction task is to predict the demand for the bikes (column cnt) given the other
features: ignore the columns instant and dteday. Use the day.csv file from the data folder.
(a) (4 points) Write a Python file to load day.csv.
2 Compute the correlation coefficient of each
feature with the response (i.e., cnt). Include a table with the correlation coefficient of each
feature with the response. Which features are positively correlated (i.e., have positive correlation coefficient) with the response? Which feature has the highest positive correlation with
the response?
(b) (2 points) Were you able to find any features with a negative correlation coefficient with the
response? If not, can you think of a feature that is not provided in the dataset but may have
a negative correlation coefficient with the response?
(c) (5 points) Now, divide the data into training and test sets with the training set having about
70 percent of the data. Import train_test_split from sklearn to perform this operation.
Use an existing package to train a multiple linear regression model on the training set using all
the features (except the ones excluded above). Report the coefficients of the linear regression
models and the following metrics on the training data: (1) RMSE metric; (2) R2 metric.
[Hint: You may find the libraries sklearn.linear_model.LinearRegression useful.]
(d) (2 points) Next, use the test set that was generated in the earlier step. Evaluate the trained
model in step (c) on the testing set. Report the RMSE and R2 metrics on the testing set.
(e) (2 points) Interpret the results in your own words. Which features contribute mostly to the
linear regression model? Is the model fitting the data well? How large is the model error?
1
https://www.kaggle.com/datasets/marklvl/bike-sharing-dataset?search=bike+demand+Washington&
select=Readme.txt. You can also find a Readme.txt file that explains all the features in the dataset.
2Refer to https://docs.python.org/3/library/csv.html on how to load a csv file in Python.
Problem 5 (10 points)
This question should be answered using the Diabetes data set that is readily available in the
Scikit-learn library.3 This data set has information about 442 patients and whether they have
suffered from diabetes or not.
(a) (2 points) Fit a multiple regression model to predict Diabetes using Age, Sex, BMI, and BP.
(b) (4 points) Provide an interpretation of each coefficient in the model. Be careful—some of the
variables in the model are qualitative!
(c) (2 points) Write out the model in equation form, being careful to handle the qualitative
variables properly.
(d) (2 points) Using the model from (c), obtain 95% confidence intervals for the coefficient(s).
Problem 6 (20 points)
We will now perform cross-validation on a simulated data set.
(a) (2 points) Generate a simulated data set as follows:
numpy.random.seed(12345)
x = numpy.random.normal(0, 1, (200))
y = x + 2 * x**2 – 2 * x**3 + numpy.random.normal(0, 1, (200))
In this data set, what is n and what is p? Write out the model used to generate the data in
equation form.
(b) (2 points) Create a scatterplot of X against Y . Comment on what you find. (Hint: You may
find matplotlib.pyplot.plot() helpful)
(c) (9 points) Set a random seed 123, and then compute the leave-one-out cross validation errors
that result from fitting the following five models using least squares:
(i) Y = β0 + β1X + ε
(ii) Y = β0 + β1X + β2X2 + ε
(iii) Y = β0 + β1X + β2X2 + β3X3 + ε
(iv) Y = β0 + β1X + β2X2 + β3X3 + β4X4 + ε
3
https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_diabetes.html. You can
find the description of this data set at https://scikit-learn.org/stable/datasets/toy_dataset.html.
(v) Y = β0 + β1X + β2X2 + β3X3 + β4X4 + β5X5 + ε
[Hint: You may find LeaveOneOut() and cross_val_score() in sklearn.model_selection
helpful.]
(d) (2 points) Repeat (c) using another random seed 12345, and report your results. Are your
results the same as what you got in (c)? Why?
(e) (5 points) Which of the models in (c) had the smallest leave-one-out cross validation error?
Is this what you expected? Explain your answer.