Dear Reader,

Thank you for your interest in the Machine Learning Interview and for
purchasing this book.

First, congratulations on picking data science as your future career! There

hasn’t been a better time to get into this field with the demand for data
science professionals far outstripping the supply.

We understand how difficult landing that first data science role can be. The
biggest hurdle can often by clearing the interviews! You might have studied
up on all the techniques and tools, brushed up on your previous experience,
and yet you can’t quite crack the interview process.

That’s why we combined our collective experience of conducting hundreds

of interviews and created the interview Guide.

This book is a sample of the far more comprehensive handbook of data

science interview questions. We have picked out some of the most intriguing
questions from that book for your perusal here.

The complete handbook is part of the ‘Ace Data Science Interviews’

course. It contains over 240 questions based on data science concepts.
There is no other resource quite like it anywhere.

Wondering what else to expect in the course? We have put together a

comprehensive 7-step framework to help you land your first data science
role. We have spoken more about this framework in an article complete with
an illustrated infographic detailing each step.

We hope you find this e-book useful and we’ll see you in the course!

You can also check out our Ace Data Science Interviews website and
podcast for more tips and tricks on how to crack the interview process.

Thanks,
Analytics Vidhya Team
 Table of Contents
1. Puzzles
2. Guesstimate
3. Case study
4. Python for Data Science
5. Statistics and Probability
6. Machine Learning
 Puzzles

Question 1

You pull out 2 balls, one after another, from a bag which has 20 blue and 13
red balls in total. If the balls are of similar colour, then the balls are replaced
with a blue ball, however, if the balls are of different colours, then a red ball
is used to replace them.

Once the balls are taken out of the bag, they are not placed back in the bag,
and thus the number of balls keep reducing. Determine the colour of last ball
left in the bag.

Answer 1
If we pull out 2 red balls, we need to replace them with a blue ball. On the
other hand, if we draw one red and one blue, it is replaced with a red one.
This implies that the red ball would always be in odd numbers, whether we
remove 2 together, or remove 1 while adding 1.

This also indicates that the last ball to stay in the bag would be a red one.
Question 2

There are 3 mislabeled jars, with apple and oranges in the first and second
jar respectively. The third jar contains a mixture of apples and oranges. You
can pick as many fruits as required to precisely label each jar. Determine the
minimum number of fruits to be picked up in the process of labeling the jars.

Answer 2
A noticeable aspect in this puzzles is the fact that there’s a circular
misplacement, which implies if apple is wrongly labelled as Apple, Apple
can’t be labelled as Orange,
i.e., it has to be labeled as A+O. We are acquainted with the fact that
everything is wrongly placed, which means A+O jar contains either Apple or
Orange (but not both).

The candidate picks one fruit from A+O, and let’s assume he gets an apple.
He labels the jar as apple, however, jar labelled Apple can’t have A+O. Thus,
the third jar left in the process should be labelled A+O.

Question 3
There are 5 pirates in a ship. Pirates have hierarchy C1, C2, C3, C4 and C5.
C1 designation is the highest and C5 is the lowest. These pirates have three
characteristics: a. Every pirate is so greedy that he can even take lives to
make more money. b. Every pirate desperately wants to stay alive. c. They
are all very intelligent.

There are total 100 gold coins on the ship. The person with the highest
designation on the deck is expected to make the distribution. If the majority
on the deck does not agree to the distribution proposed, the highest
designation pirate will be thrown out of the ship (or simply killed). Only the
person with the highest designation can be killed at any moment. What is the
right distribution of the coins proposed by the captain so that he is not killed
and does make maximum amount?
Answer 3
The solution of this problem lies in thinking through what will happen if all
the pirates were thrown one by one and then thinking in reverse order.

Let us name pirates as A,B,C,D and E in hierarchy (A being highest).

If only D and E are left at end, D will simply give 0 coins to E and still escape
because majority cannot be reached. Hence, even if E gets 1 coin he will
give his vote to the distributor.

If C, D and E are there on the deck, C will simply give one coin to E to get
his vote. And D simply gets nothing. Hence, even if D gets 1 coin he will give
his vote to the distributor.

If B,C,D and E are there on the deck, B will simply give one coin to D to get
his vote. C & E simply gets nothing.

If A,B,C,D and E are there on the deck, A simply gives 1 coin each to C and
E to get their votes.

Hence, in the final solution A gets 98 coins and only C & E get 1 coin each.
 Guesstimates

Question 4
Estimate the number of cigarettes consumed monthly in India.

Answer 4
The population of India, i.e., 1.2 billion. Following is an effective way to
segment this population:

Following were the key considerations in building the segmentation and the
intermediate guesses:

The rural population consumes far lesser cigarettes than urban because of
the purchasing power difference.

Male consume more cigarettes than female in both urban and rural
populations.
Children below 16 years consume a negligible number of cigarettes.

Male to Female ratio in Urban is closer to 1 than that of Rural.

Male to Female ratio in younger generations is closer to 1 than that of older.

This is because of the increase in awareness level.

Bulk of population start smoking after getting into a job and hence the
average number cigarettes are higher in older groups.

Total number of cigarettes from the supply side also come to around 10
Trillion, which gives a good sense check on the final number.

Question 5
How many iPhones are currently being used in China?

Answer 5
Identify the variables to apply to this problem.

Population of China: Approximately 1.4 billion people.

There are several different approaches from this point; one approach is to
make assumptions around the number of people that can afford iPhones
rather than considering the number of households.

Based on very basic knowledge of China, even though the country is

experiencing extraordinary economic growth, you might assume that the
majority of the population is still very low-income and cannot afford an
iPhone. Thus, you might estimate that 20% of the population could afford an
iPhone.

Therefore, the total potential market size is 20% × 1.4 billion = 280 million
iPhones.
What percent of this total market size is penetrated? There are many
competing products that are cheaper, therefore we estimate that 20% of this
segment is currently using an iPhone.

Using these estimates, 20% × 280 million = 56 million iPhones are currently
being used in China.
 Case study

Question 6
There are multiple cab services these days, and you, as a customer has to
make an efficient decision, and select the cab based on your requirement.
Some important terms that we will use throughout are defined below:

1. Base Fare: Initial amount billed to sit in a cab

2. Excess km fare: Billed amount on distance after complimentary ride
3. Time fare: Billing on the time spent in the cab
4. Minimum fare: This is the minimum amount you will be billed
5. Tolls and excess fee: This is the excess charge you need to pay to
compensate for the long distances outside main city
6. Taxes: Taxes is are over and above the bill
7. Premium multiplier: In crowded/congested time, you will be bill
something like 1.4 – 2.5 X of the actual bill amount.

Here are the details of the three cab services that you must compare.
1) The first cab service, let’s call it A, has the following fare for the three
types of cabs - micro, mini, prime.
 In addition to this, minimum fare of Micro is fixed at Rs. 50.

2) The second service, called B, also has three types of cabs - nano
(equivalent to micro), tata indica (equivalent to mini), sedan
(equivalent to prime), and here are the prices of the same.

3) The third company, company C, has the following fare for the cabs
offered. They do not have a micro or nano but on the other hand
provide an XL.

You need to suggest the most optimised cab, for the following situations-

1. Which of the MICRO vehicles will be cheapest if your distance lies

between 1 to 8 kms?
2. Which MINI vehicles are the cheapest if your distance is between 1 to
10 kms?
 3. If you get a free upgrade from Company A - Micro to Mini, will it be
cheaper than Company C, Mini for distance 2-6 kms?
4. Company C is charging a multiplier of 2.1 and Company A is charging
a multiplier of 1.4 on their Sedan Vehicles (Company A Prime vs.
Company C X). Which one will cost less ?
5. You have already booked UberGo for a multiplier of 1.5 and now you
are getting a Company A Mini vehicle without peak charges? The
challenge is that if you cancel an Company C, you will incur a
cancellation charge penalty. However, if you choose to cancel, you
stand a chance to save on peak charges. At what distance will you
break even on the cancellation charges on Company C, in case you
choose to go ahead with Company A?

Answer 6
We shall take up the questions one by one.

1. Since only company A and B offer the micro cabs, we will choose a
cab from one of the two. Since the distance is provided to be 1-8 km
range, let’s check out the price for this distance for both the company.

If the distance is less

than 4, a cab from
company A should be booked while for distance more than 4, cab
from company B should be preferred.

2. Since all the companies have an option of a cab type mini, so we

must check the cab fare for all the three companies.
 We clearly see that Company C is the cheapest car for the distance
range and also seem to be the cheapest option for distances beyond
10 km looking at the trend.

3. We will compare company A micro and company C mini. The trend

looks to be very interesting. Company C mini starts at a slightly lower
rate than Company A Micro but Company A takes over after 5 km
distance. So, again our answer will be cannot be determined in the
distance range. However, it is interesting to notice that a Company A
Micro taxi comes out to be more expensive than Company C Mini car
for shorter than 5 km

distance.
4. Multipliers are often added in peak traffic hours. It becomes very
difficult to compare rates in such cases. Here’s one of those
scenarios. Company A is generally more expensive compared to
Company C, but in such extreme multiplier cases, we see that
Company A comes out to be cheaper option throughout.

5. What you need to keep in mind is the cancellation charges of

Company C. Here is the table for Company A vs. Company C :
As it can be seen from both table and the graph, the break-even only
happens between 13-14 kms. Hence, you should make a switch only
if you want to travel more than 13 kms
 Python for data science

Question 7
The two most commonly used data structures in python are lists and
dictionaries. Can you list down the basic differences between list and
dictionary.

Answer 7
Lists store values, in an ordered sequence. Each element is numbered,
starting from zero, i.e. the first element in a list is numbered 0, the second
1, the third 2, and so on. We can remove values from the list, and add new
values to the end.
Example: list1 = [1,2,3,4,10,100]

Dictionary is an unordered set of key: value pairs. Unlike lists, instead of

using numbers to represent values, we have keys in a dictionary. The
values in the dictionary can be removed/ modified and new values can be
added.
Example: dict={‘English’: 90, ‘Mathematics’: 95, ‘Physics’:80}

Question 8
Suppose I have a list list1=['a', 'b', 1, {'Name': 'ABC', 'age': 22}], what will be
the result of the following command?
>>list1[1]
>>list1[3]
>>list1[3]['age']

Answer 8
- The list has index starting from 0, so the output of the command list1[1]
will ba ‘b’.
- Since the fourth element in the list is a dictionary, the result would be
{'Name': ‘ABC’, 'age': 22}
- List1[3]['age'] will give the output 22

Question 9
In order to add new elements to a list, we can use one of the two
commands - list.append() or list.extend()? For a given list: a=[1,2,3,4], what
will be the result of the following commands?
>>a.append([6,7,8])
>>a.extend([6,7,8])

Answer 9
Append is used to add a single value at the end of the list. For the given
example, append assumes [6,7,8], to be a list. So the result will be =
[1,2,3,4,[6,7,8]]

Extend is used to add multiple elements to the list. In this case, the
numbers 6,7 and are considered to be three different elements. The output
of second command will be = [1,2,3,4,6,7,8]

Question 10
What will be the output of the following?

➢ for i in range(2.0):
print(i)
Answer In python range function can only iterate using integer values.
Since 2.0 is a floating point number thus the above code will result in
an error.

➢ for i in range(2,10):
print(i)
Answer The above code prints numbers from 2 to 9.

➢ for i in range(2,10,2):
print(i)
Above code prints all the even numbers from 2 to 8(10 is exclusive),
because the third parameter in range is the step-size for iteration.

Question 11
What is the difference between the following code lines?

>> df.drop([‘Age’], axis = 1)

>> df.drop([‘Age’], axis = 1, inplace=True)

Answer 11
The first case, since the value for parameter inplace is not specified, the
default inplace=False is considered. It prints the dataframe without the
column specified in the parameter, ‘Age’ in this case. But this does not
make any change to the original dataframe.

When the parameter inplace is set to be true, the column is dropped from
the original dataframe.

Question 12
What is the difference between file.tsv and file.csv? How are they used?

Answer 12

• .tsv and .csv are two different file formats referring to tab separated
values and comma separated values respectively.
• .tsv are often used in text data as comma can represent a
punctuation.
• .tsv and csv files can be loaded into a pandas dataframe using the
read_csv function.
Question 13
What are lambda functions, when to use them and when not to use them.?

Answer 13

Lambda functions are anonymous functions and are not bound to any
name/identifier. They are used when we can to perform a “small set of
actions” and which does not require defining a separate function. They are
lightweight and efficient. They are not recommended to use when the set of
actions is not “small” or are too complex. Defining a separate function is
better in this scenario.

Question 14

Which is more efficient, Iteration or recursion?

Answer 14

Python is not primarily a functional language and therefore there is a little

overhead when using recursions, Therefore iterative methods are mostly
efficient. But, there are problem where applying iteration can be tedious (
Eg: Tower of Hanoi ) and therefore can be efficiently solved using the
recursion techniques.

Question 15

What do you mean by scope of a variable?

Answer 15
A scope of a variable refers to the environment to which it is restricted to, A
variable of higher scope can used used at lower scope but reverse is not
true.
 Statistics and Probability

Question 16
The mean of a distribution is 25, the median is 23, and the mode is 20.
Based on the given information can you determine whether the plot for this
distribution, will be positively skewed or negatively skewed?

Answer 16
Since the mode value is less that the mean and median, the data will be
positively skewed. Here is a diagrammatic representation for the same.
Question 17
The interquartile range is the difference between the minimum and
maximum value of the numbers in a given series. Is the statement correct?

Answer 17
No, the difference between the minimum and maximum value is the range
of the series. The interquartile range is the difference between the the first
quartile (value below which the lower 25% of the data exist) and the third
quartile (value below which the lower 75% of the data exist).

Question 18
How many values, in the below series, fall within one standard deviation of
the mean?
180, 313, 101, 255, 202, 198, 109, 183, 181, 113, 171, 165, 318, 145, 131,
145, 226, 113, 268, 108

Answer 18
On calculation, the mean and the standard deviation of the given data
comes out to be 181.25 and 64.65 respectively. Now, one standard
deviation above an below the mean would give the range 117 to 245. Total
11 numbers fall in this range.

Question 19
Two unbiased coins are tossed. What is the probability of getting at most
one head?

Answer 19
The probability is ¾. Since the total possible outcomes when two coins are
tossed will be 4 (HH, HT, TH, TT). We are to find at most 1 head, so the
favorable outcomes are (TT, HT, TH). Therefore, Probability = 3⁄4.

Question 20
Normal distribution is symmetric about the origin (0,0)?
Answer 20
No, the normal distribution is symmetric about the mean. But for a standard
normal curve the values are scaled between -1 to1. In this case, the value
of mean is 0 and standard deviation is 1. So the standard normal curve is
symmetric about 0.

Question 21
What happens to the confidence interval when we introduce some outliers
to the data?

Answer 21
The confidence interval depends on the standard deviation of the data. On
introducing outliers in the data, the standard deviation increase, and so
does the confidence interval.

Question 22
What does confidence interval = 95% mean?

Answer 22
On repetitive sampling, 95% of the times, the interval estimates will contain
the population mean.

Question 23
What does the p-value for a statistical data signify?

Answer 23
The p value for a statistical test is used to draw conclusions, basically
deciding to accept of reject the null hypothesis.The range of p-value is
always between 0 and 1. Generally the threshold for p value is set to be
0.05. When the value is below 0.05, the null hypothesis is rejected. If the
value is equal to or greater than the threshold, 0.05 in this case, the null
hypothesis is not rejected.
Question 24
Consider the following two situations -

Situation 1: The doctor identifies a male patient as pregnant.

Situation 2: The doctor identifies a female pregnant lady as not pregnant.

Given that the null hypothesis is that the patient is not pregnant, classify the
above two situations as Type I and Type II error.

Answer 24
For the first situation, the null hypothesis is incorrectly rejected, which is a
type II error. In the second situation, the null hypothesis is falsely accepted
hence it is type I error.
 Machine learning

Question 25
The linear regression model fits a line in order to model the relationship
between independent and dependent variables. So how does the model
decide the right line?

Answer 25

➢ As in the figure above, there could be multiple lines fit on the

data points. The right line is decided such that the error value is
minimum (as compared to other lines).
➢ Selecting the best fit line is equivalent to selecting the right set
of coefficients for the line. This is an optimization problem.
➢ To select the best fit line, we use gradient descent optimization
technique.
 ➢ Initially the coefficients for independent variables are randomly
assigned and the error is calculated. These weights are then
continuously updated to reduce the error.

(if you are not familiar with the concept of gradient descent algorithm, you
must read this article: Introduction to Gradient Descent Algorithm)

Question 26
If you are given the two variables V1 and V2 such that
1. If V1 increases then V2 also increases
2. If V1 decreases then V2 behavior is unknown
What is the value of pearson’s correlation between V1 and V2?

Answer 26
We cannot comment on the correlation coefficient by using only statement
1. We need to consider the both of these two statements. Consider V1 as
x and V2 as |x|. The correlation coefficient would not be close to 1 in such a
case.

Question 27
Suppose we have a classification problem where the target variables have
4 classes - A, B, C, D. How can we use logistic regression for this problem
statement?

Answer 27
The logistic regression can be used for multiclass classification by
implementing the “one-versus-rest” technique where we take up each class
and try to classify whether the data points belong to that particular class or
not.

Suppose we have four classes A, B, C and D. We can start with classifying

the data points as A vs B/C/D (not A), and then for class B vs not B, so on.

Model-> Mode l1 Model 2 Model 3 Model 4

(A vs B ,C (B vs A, C, D) (C vs A, B, D) (D vs A, B,
 ,D) Calculating Calculating C)
Calculating probability of probability of Calculating
probability outcome B outcome C probability of
of outcome outcome D
A

Question 28
Suppose I applied a logistic regression model on data and got training
accuracy X and testing accuracy Y. Now I add 2 new features in data.
Select option(s) which are correct in such case.

Note: Consider remaining parameters are same.

1. Training accuracy always decreases.

2. Training accuracy always increases or remain same.
3. Testing accuracy always decreases
4. Testing accuracy always increases or remain same

Answer 28
Training accuracy will increase but testing accuracy would increases only
when the new variables are found to be significant.
Question 29
Suppose we have 2000 data points and 20 features. What is the maximum
number of leaf nodes possible with a decision tree? Also, can you calculate
the number of levels to which the tree will grow (depth of tree).

Answer 29
The maximum number of leaf nodes will be equal to the number of
datapoints in the dataset (such that each leaf node has only one
sample/datapoint). In this case the maximum number of leaf nodes will be
20,000.

Question 30
If we let the tree grow to the maximum length, there are chances of
overfitting. How will you make sure that your decision tree model does not
overfit?

Answer 30
If we let the tree grow to the maximum length, there are high chances that
the model overfits on the training data. So we should reduce the size of the
tree, which is called tree pruning. To do so, we can set certain conditions or
constraints on the various parameters of decision tree model - such as max
depth, min leaf sample size etc.

Consider the example of max depth, if we have a total of 500 data points,
then the tree grows till each leaf node has only one sample. Instead, if we
set a constraint on the parameter max_depth = 3, the tree would only grow
for three levels.
Question 31
A random forest model is an ensemble of multiple decision trees. Consider
the following two scenarios -

Case 1: Build a random forest with 10 estimator (i.e. 10 decision trees).

Case 2: Creating 10 decision trees and averaging the predictions from
these trees to calculate the final predictions.

Will both the models give the same results? (considering the dataset and
other parameters are same in both the cases)

Answer 31
When training a decision tree classifier, splits are done using all of the data
points and all of the features whereas in case of random forests every
estimator(decision tree) is based only on a random subset of the
dataset(thus the name random forest).

Hence, in the case of multiple decision trees trained on the same dataset ,
all the trees will be same along with their predictions for a given data point
whereas because of change in distribution of data for different decision
trees in case of random forests, the predictions can be different for the
same data point from different estimators.Thus Case 1 and Case 2 will
have different results.

Question 32
Two most important ensemble learning methods are bagging and boosting.
Can you explain the difference between bagging and boosting?

Answer 32
Bagging (or Bootstrap Aggregating): In this technique, multiple subsets of
the complete data are created and a model is trained on each of these
subsets. The final prediction is made by combining the values predicted
from each of these models trained on different subsets. Various algorithms
which use bagging technique are Bagging estimator, Random Forest, Extra
trees.

Boosting: It is a sequential process where each subsequent model

attempts to correct the errors from the previous model. This is done by
giving higher weights to the observations which were incorrectly predicted.
Final model (strong learner) is the weighted mean of all the models (weak
learners). AdaBoost GBM, XGBoost, etc are some of the algorithms which
use boosting technique.

Question 33

What does the K in KNN mean? How do you find the optimal value for K?

Answer 33
K is a hyperparameter for the KNN algorithm which decides the number of
training data points that are used for predicting the label for the data point
during the testing phase. For finding the optimal value of k , we can try all
possible values of K from 3 upto a specific value and monitor the loss for
the validation data. Another possible solution is the elbow technique where
we plot the total sum of squared error for different values of k and find the
point where the total sum of squares starts increasing.

Here k=8 seems like the optimal value for our dataset.
Question 34
While doing K means algorithm, consider the following two cases-
Case1: The cluster centers are updated after every datapoint assignment
Case2: The cluster centers are updated after assigning every data point to
some cluster.

We have applied k means on our dataset to calculate the cluster centers,

now one feature of the dataset was scaled by some factor M, do the cluster
centers change? What if all the features were scaled?

Answer 34
In case only one feature was scaled , the cluster centers will need to be
updated as it will increase/decrease the value of single feature that will lead
to the change of cluster centers thus there might be some points whose
clusters will get changed.But in case all the features were scaled , the
cluster centers will also get scaled by the same factor but in this case the
clusters will not change.

Question 35
Suppose we were given a dataset that consists of all numerical features,
we applied linear regression but the results were not satisfactory. Further
we also used SVM but still the performance did not improve , what seems
to be the problem? Find out the best decision boundary from the ones
given below

Answer 35
If both linear regression and SVM are not giving satisfactory results on a
dataset, this implies that the classes are not linearly separable. In such a
situation we can use the kernel trick of SVMs to project the dataset into
higher number of dimensions such that the classes become linearly
separable.

The best decision boundary is the one that maximizes the margin on both
sides.(SVM theory)

Question 36
Both being tree based algorithm, how is random forest different from
Gradient boosting algorithm (GBM)?

Answer 36
The major difference between random forest and gradient boosted trees
lies in the fact that the individual estimators in random forests are
independent of each other whereas in case of gradient boosted trees the
individual estimators are not independent. The estimators in random forests
are built parallely whereas in GBTs the estimators are built incrementally.
Also the data points for which the current estimator was not able to perform
well are given more preference in the next estimators so that the whole
model is able to generalize well over the complete dataset.

Question 37
Which of the following are mandatory data pre-processing step for XGB?
1. Impute Missing Values
2. Remove Outliers
3. Convert data to numeric array / sparse matrix
4. Input variable must have normal distribution
5. Select the sample of records for each tree/ estimators

Answer 37
Converting data to numeric array/sparse matrix is a mandatory pre-
processing step for XGBOOST as xgboost cannot handle categorical
features.
Question 38
The features of our dataset show strong correlation , is Naive Bayes a good
choice for such a dataset? How will you modify the dataset such that Naive
Bayes shows good results?

Answer 38
Strong correlation between features implies dependence. Thus Naive
Bayes is not a good choice for such a dataset because Naive Bayes
assumes that the features of the given dataset are independent of each
other. In such a case we can apply PCA to identify the independent
dimensions from the dataset and then use Naive Bayes.

Question 39
The categorical variables can further be classified as cardinal and ordinal.
Can you explain the difference between them and how can we deal with
these variables?

Answer 39
The variables which have a certain number of categories fall under the term
‘Cardinal variables’. These variables do not have an order between them.
For example, the variable ‘Gender’ can have two categories - Male and
Female, and there is no order to these categories. Most machine learning
models cannot deal with categorical variables. So we use one hot encoding
technique to convert these categorical variables into binary.

While Ordinal variables have categories which contain an order between

them, such as ‘class of a student’ which can be I, II,....IX, X, and so on. We
know that IV and V are higher than I and II. So we can say that these
categories have an order. Although one hot encoding technique can be
applied on these variables, it is preferred to use label encoding (since the
order in the variables is important)

Question 40
Have built two models “Model 1” and “Model 2”. "Model 1" has high bias
and low variance where as “Model 2” has low bias and high variance.
Which of these two models have higher chance of overfitting or under-
fitting?

Answer 40
A model with high bias and low variance tends to overfit on certain feature
or data points. The model will not be generalized and performs poorly, thus
has a higher chance of underfitting. On the other hand, a model with high
variance and low bias will have a higher chance of overfitting. A simple
diagram (shown below) can be used to verify the above statements.

Question 41
There are multiple machine learning algorithms, such as linear regression,
logistic regression, random forest, xgb etc. How will you select which model
to implement?

Answer 41
The model to be used is often decided based on the dataset we need to
work on.

➢ For a regression problem, we can choose linear regression,

random forest regressor etc.
 ➢ Similarly for classification problem, the choice of models would
be slightly different.
➢ For imbalanced dataset, boosting techniques work really well.
➢ In case the dataset has a very high number of categorical
variables, we can use CatBoost.
➢ For very large datasets, light GBM has proved to work better
than other algorithms.

Question 42
What do you understand by Type I vs Type II error ?

Answer 42
Type I error is committed when the null hypothesis is true and we reject it,
also known as a ‘False Positive’. Type II error is committed when the null
hypothesis is false and we accept it, also known as ‘False Negative’.

In the context of confusion matrix, we can say Type I error occurs when we
classify a value as positive (1) when it is actually negative (0). Type II error
occurs when we classify a value as negative (0) when it is actually
positive(1).
Question 43
How can you differentiate or classify a problem statement as time series
problem? For instance, consider the following two problem statements;
which one would you apply a time series model on?

Problem 1- Based on - the income per month, date of issuing loan, and the
loan type, we have to predict the loan amount for the person.

Problem 2- The target is to predict the hourly temperature using the for a
particular region given the past data for the same.
Answer 43
The data points in a time series dataset are equally spaced and the target
variable for each row is dependent on the past dependent and independent
variables.

From the given two problems, the second problem represents a time series
since the next hour’s temperature depends on the previous value. While in
the first problem, the loan amount for an individual person does not depend
on the loan taken by the previous person.
Question 44
Identify the p, q values from the following acf pacf plots-

Answer 44
The value of p and q are taken from as the point where the dotted line is
intersected the first time. ACF plot is used to determine the value of q while
PACF plot is used to find the parameter p.

In the ACF plot shown above, the first intersection is at the value one, so
the value of q must be taken as 1. Similarly, the value of p is also 1.
Question 45

Suppose we have the following architecture for a Multi-Layer Perceptron a.

Number of nodes at input layer = 10 b. Nodes at hidden layer = 5 c.
Number of output node= 1 What are the total number of connections?

Answer 45

Each node in the input layer has a connection to each node in the hidden
layer. So this makes the total number of connections between input and
hidden = 50. Then we will have 5 connections between the hidden and
output layer. So total number of connections in the above architecture is
55.

Question 46

There are many different activation functions like relu, sigmoid, softmax or
tanh. How can one decide which activation function to use? How will you
decide the activation functions for input and output layer?

Answer 46

Sigmoid is an activation function that returns a value between 0 and 1, it is

primarily used for the output node of the Binary Classification, but can also
be used in the hidden layers.

Tanh is an activation function that returns a value between -1 and 1, it is

primarily used in hidden layers of Neural Networks.

Relu is an activation function that returns a linear value when the input is
positive and returns 0 when the input is 0 or negative. It is primarily used in
hidden layers of Neural Networks.

Softmax is an activation function which is primarily used in the last/output

layer of multi-class classification problem.
Question 47
Dropout and DropConnect are both regularization techniques for Neural
Network. Is there a difference between these two? How is setting dropout
=0.3 different from drop connect =0.3?

Answer 47

Drop-Out

The function drop-out in a layer assigns a probability p to every node in that

layer such that, that node will not be included in the computation during the
runtime with respect to probability p. (0.3 in question)

Drop-Connect

The function drop-connection in a layer is a probability p for every node in

that layer such that, there is a chance p such that , that node will skip a
connection to consecutive layer by the probability p.
Question 48

During feature selection - we remove features thus giving less information

to the model for training. Will this affect performance of the model?

Answer 48

Feature selection is selecting the most relevant attributes for the predictive
model. Only the redundant features are removed from the model so that it
does not adversely affect the accuracy of the model. Having fewer but
relevant features reduces the model complexity and training time.

Question 49

There are broadly two common methods to convert categorical variable to

number, Label Encoding and One Hot Encoding. How do you decide which
method to use when?

Answer 49
Most machine learning models cannot deal with categorical variables and
hence we need to perform one hot encoding or label encoding before
feeding the data to the model.

The variables which have a certain number of categories fall under the term
‘Categorical variables’. These variables do not have an order between
them. For example, the variable ‘Gender’ can have two categories - Male
and Female, and there is no order to these categories. In this case, we can
use the one hot encoding technique to convert the categorical columns into
two columns with binary values.

Question 50

You came to know that your model is suffering from low bias and high
variance. Which
algorithm should you use to tackle it? Why?

Answer 50
Low bias occurs when the model’s predicted values are near to actual
values. In other words, the model becomes flexible enough to mimic the
training data distribution. While it sounds like great achievement, but not to
forget, a flexible model has no generalization capabilities. It means, when
this model is tested on an unseen data, it gives disappointing results. In
such situations, we can use bagging algorithm (like random forest) to tackle
high variance problem. Bagging algorithms divides a data set into subsets
made with repeated randomized sampling.