Total No. of Pages Roll No.

……
FIFTH SEMESTER [Link]. (CSE)
MID SEMESTER EXAMINATION September-2022
CO327 MACHINE LEARNING
Time: 1:30 Hours Max. Marks: 20
Note: Answer ALL questions.
Assume suitable missing data, if any.
CO# is course outcome(s) related to the question.
L# is the cognitive level required to solve the question.
1[a] A tollbooth collects the data of various cars passing through it. The
following attributes are recorded: speed of the car, gender of the driver,
time of arrival, car registration number, age bracket of the driver (young,
middle, old), number of co-passengers, and driving license number. The
booth operators want to design a machine learning (ML) model to
predict the speed of cars using this model. Identify the features that can
be used for ML model design. Also, identify whether it is a classification
problem, regression problem, or none. [Give one-line justification (not
more) for each selection/answer] [1+1] [CO1] [L2]
[b] The probability distribution f(X) of a random variable X is given in
Table. I. Compute the mean and variance of X. [1+1] [CO2] [L3]
Table. I
X 0 1 2 3
f(X) 1/7 3/7 2/7 1/7

2[a] A travel agency wants an automated system to predict travel costs. The
agency has the following data available with it.
Table II
S. No. Distance Travelling Cost
(in Km) (in Rupees)
1 1 2.75
2 2 3.5
3 3 4.25
4 4 5
5 5 5.75
Page 1 of 3
 Formulate the above problem as a linear model h(x) = w0 + w1x to predict
the travelling cost for a given distance. The parameter w0 is 2 (optimal).
Apply gradient descent algorithm to find optimal parameter w1. The
learning rate for the first epoch is 0.073, and for the second epoch and
later, the learning rate is 0.091. Let the initial value of w1 is 0.5.
[4] [CO1, CO2] [L3]
[b] In logistic regression, binary cross-entropy is used as the cost function
for two-class classification. Illustrate (considering one sample) that the
cost function will have a single optimum so that the gradient descent
algorithm converges to the global optima. [3] [CO2] [L4]

3[a] A factory is producing papers. The quality control unit applies two types
of testing (durability test and strength test) to assess paper quality. The
data for the same is given below:
Table III
S. No. 1 2 3 4 5 6 7 8
Durability 7 6 7 6 3 1 4 3
Strength 7 4 4 5 4 4 3 5
Quality Good Bad Good Good Bad Bad Bad Bad

In general, the factory produces 720 good quality papers out of 1000.
Use k-nearest neighbor (KNN) with k = 1, and 3 to predict the quality
of a new paper (durability = 5, strength = 5). [2+1] [CO3] [L3]

[b] Now, suppose (in above question 3[a]), we define some distance-based
probabilistic classifier instead of KNN. The likelihood of belonging to
a class for a new sample is 1⁄𝑑 . Here d is the Euclidian distance of a
new sample from nearby samples of the same class. If there are multiple
neighbouring samples of a class, the overall likelihood is calculated by
the union of all likelihoods. Assume a cutoff distance dcf; beyond that,
no sample is considered in calculating overall likelihood. [Hint: p(A∪B)
= p(A) + p(B) - p(A∩B)]
Consider dcf is the maximum distance of new sample (durability =
5, strength = 5) from other samples in 3[a] for KNN with k = 3. Predict
the quality of a new paper (durability = 5, strength = 5) using posterior
probability. Also, compare the performance of this probabilistic
classifier with KNN {Maximum two sentences}. [2+1] [CO3] [L3, L5]

Page 2 of 3
4 A career counselling agency wants an automated system to advise for
MS programs. It has previous data (given in Table IV) of students who
have succeeded or failed in MS programs. The data contains two
attributes of each student: CGPA (High, Medium, Low) and whether or
not they have published a good research paper (Yes, No). An ML
engineer is hired to develop such a system. He thought of applying a
decision-tree algorithm but wanted a new criterion of data division (in
subsets). He got an idea for the same, inspired by the F1 score. In the F1
score, he replaced precision and recall with the two classes (succeed,
failed) probabilities and named it the G1 score. Apply this newly defined
G1 score and develop a full decision tree. [Use the weighted average of
G1 scores of subsets to compare with the G1 of the original set (before
division)]. [3] [CO1, CO2] [L3]
Table IV
S. No. CGPA Publication Result (MS)
1 Low No Failed
2 Low Yes Succeed
3 Medium No Failed
4 Medium Yes Succeed
5 High No Succeed
6 High Yes Succeed

---Best of Luck---

Page 3 of 3