Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Python Data Science Interview

Questions
Every data science interview has many Python-related questions, so if you
really want to crack your next data science interview, you need to master
Python.

Q.1 What is a lambda expression in Python?

Ans. With the help of lambda expression, you can create an anonymous
function. Unlike conventional functions, lambda functions occupy a single line
of code. The basic syntax of a lambda function is –
lambda arguments: expression
An example of lambda function in Python data science is –

x = lambda a : a * 5
print(x(5))
We obtain the output of 25.
Q.2 How will you measure the Euclidean distance between the two arrays in
numpy?
Ans. In order to measure the Euclidean distance between the two arrays, we
will first initialize our two arrays, then we will use the [Link]() function
provided by the numpy library. Here, numpy is imported as np.
a = [Link]([1,2,3,4,5])
b = [Link]([6,7,8,9,10])
# Solution
e_dist = [Link](a-b)
e_dist
11.180339887498949

With data integrity, we can define the accuracy as well as the consistency of
the data. This integrity is to be ensured over the entire life-cycle.

Q.3 How will you create an identity matrix using numpy?

Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Ans. In order to create the identity matrix with numpy, we will use the
identity() function. Numpy is imported as np
[Link](3)
We will obtain the output as –

array([[1., 0., 0.],

[0., 1., 0.],
[0., 0., 1.]])
Q.4 You had mentioned Python as one of the tools for solving data science
problems, can you tell me the various libraries of Python that are used in Data
Science?
Ans. Some of the important libraries of Python that are used in Data Science
are –
 Numpy
 SciPy
 Pandas
 Matplotlib
 Keras
 TensorFlow
 Scikit-learn
To crack your next Data Science Interview, you need to learn these top Python
Libraries now.
Q.5 How do you create a 1-D array in numpy?
Ans. You can create a 1-D array in numpy as follows:
x = [Link]([1,2,3,4])

Where numpy is imported as np

Q.6 What function of numpy will you use to find maximum value from each row
in a 2D numpy array?
Ans. In order to find the maximum value from each row in a 2D numpy array,
we will use the amax() function as follows –
[Link](input, axis=1)

Where numpy is imported as np and input is the input array.

Q.7 Given two lists [1,2,3,4,5] and [6,7,8], you have to merge the list into a single
dimension. How will you achieve this?
Ans. In order to merge the two lists into a single list, we will concatenate the
two lists as follows –
list1 + list2

We will obtain the output as – [1, 2, 3, 4, 5, 6, 7, 8]

Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Q.8 How will you create an identity matrix using numpy?

Ans. In order to create the identity matrix with numpy, we will use the
identity() function. Numpy is imported as np
[Link](3)

We will obtain the output as –

array([[1., 0., 0.],

[0., 1., 0.],
[0., 0., 1.]])
Q.9 How to add a border that is filled with 0s around an existing array?
Ans. In order to add a border to an array that is filled with 0s, we first make an
array Z and initialize it with zeroes. We first import numpy as np.
Z = [Link]((5,5))
Then, we perform padding on it with the help of pad() function.
Z = [Link](Z, pad_width=1, mode='constant', constant_values=0)
print(Z)

Q.10 Consider a (5,6,7) shape array, what is the index (x,y,z) of the 50th element?
Ans.
print(np.unravel_index(50,(5,6,7)))

Q.11 How will you multiply a 4×3 matrix by a 3×2 matrix ?

Ans. There are two ways to do this. The first method is for the versions of
Python that are older than 3.5 –
Z = [Link]([Link]((4,3)), [Link]((3,2)))
print(Z)
array([[3., 3.],
[3., 3.],
[3., 3.],
[3., 3.]])

The second method is for Python version > 3.5,

Z = [Link]((4,3)) @ [Link]((3,2))

Note: This is not enough, a lot of NumPy related questions are asked in the Data
Science Interview. Therefore DataFlair has published Python NumPy Tutorial – An
A to Z guide that will surely help you.

Scenario-based Data Science

Interview Questions and Answers
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Below are 20 scenario or situation based interview questions provided by

experts of Data Science. If you want to take a step ahead among all the
aspirants then practice the below questions sincerely.

Q.12 Suppose that you have to train your neural networks over a dataset of 20
GB. You have a RAM of 3 GB. How will you resolve this problem of training
large data?
Ans.
 We will train our neural network with limited memory as follows:
 We first load the entire data in our numpy array.
 Then we obtain the data through passing the index to the numpy array.
 We then pass this data to our neural network and train it in small batches.
You can’t afford to miss Neural Network for data science interview preparation.
Learn it through the DataFlair’s latest guide on Neural Networks for Data Science
Interview.
Q.13 If through training all the features in the dataset, an accuracy of 100% is
obtained but with the validation set, the accuracy score is 75%. What should be
looked out for?
Ans. If the training accuracy of 100% is obtained, then a verification of
overfitting is required in our model.
Q.14 Suppose that you are training your machine learning model on the text
data. The document matrix that is created consists of more than 200K
documents. What techniques can you use to reduce the dimensions of the data?
Ans. In order to reduce the dimensions of our data, we can use any one of the
following three techniques:
 Latent Semantic Indexing
 Latent Dirichlet Allocation
 Keyword Normalization
Q.15 In a survey conducted, the average height was 164cm with a standard
deviation of 15cm. If Alex had a z-score of 1.30, what will be his height?
Ans. Using the formula, X= μ+Zσ, we determine that X = 164 + 1.30*15 =
183.5. Therefore, the height of Alex is 183.50 cm.
Q.16 While reading the file ‘[Link]’, you get the following error:
Traceback (most recent call last):
File “<input>”, line 1, in<module>
UnicodeEncodeError: ‘ascii’ codec can’t encode character.
How will you correct this error?
Ans. In order to correct this error, we will read the csv with the utf-8 encoding.
pd.read_csv(“‘[Link]”, encoding=’utf-8′)
Q.17 Assume that you have to perform clustering analysis. The first step towards
any data science problem including clustering is data cleaning. However, in this
case of clustering analysis you have a lesser number of data points. What
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

strategy would you use while performing data cleaning prior to the clustering
operation?
Ans. Capping and Flooring would be the most appropriate strategy for
performing data cleaning prior to the clustering operation.
Q.18 Assume that you are given a data science problem that involves
dimensionality reduction as a part of its pre-processing technique. You are
required to reduce the original data to k dimensions using PCA and then use
them as projections for the main features. What value of k would you select –
high or low to decrease the regularization?
Ans. In order to preserve the characteristics of our data, the value of k will be
high, therefore, leading to less regularization.
Q.19 For a given dataset, you decide to use SVM as the main classifier. You select
RBF as your kernel. What would be the optimum gamma value that would allow
you to capture the features of the dataset really well?
Ans. In SVM, the gamma parameter denotes the influence of the points that
are either near or far away from the dividing hyperplane. When the gamma is
high, the model will be able to capture the shape of the data quite well.
Master SVM concepts with DataFlairs best ever tutorial on Support Vector
Machines
Q.20 Suppose that you have to perform transformation operation on an image.
The operation is a basic rotation. The point to be rotated has the coordinates
(2,0) to a new coordinate of (0,2). How will you perform this operation?
Ans. In order to rotate the image from the point (2,0) to the point (0,2), we
will perform matrix multiplication where [2,0] will be represented as a vector
that will be multiplied with the matrix [ [0,-1] , [1,0] ]. As a result of their dot
product, we will obtain the new coordinate point of (0,2).
Q.21 Assume that while working in the field of image processing. You have to
deploy Finite Difference Filters. However, these filters are very vulnerable to
additional noise. What will you do to reduce the noise to the point of minimal
distortion?
Ans. In order to reduce the noise to the point of minimal distortion while
using the Finite-Difference Filters, we will make use of Smoothing. Smoothing
is used in image processing to reduce noise that might be present in an image
which can also be used to produce an image that is less pixelated.
Q.22 Assume that for a binary classification challenge, we have a fully connected
architecture comprising of a single hidden layer with three neurons and a single
output neuron. The structure of the input and output layer is as follows –
Input dataset: [ [0,1,1,0] , [1,1,0,0] , [1,0,0,1], [1,1,0,0] ]
Output: [ [0] , [1] , [0] ]
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

For performing model training, the weights have been initialized for both the
input and output layer as 1. Based on this, will the model be able to learn from
the patterns?
Ans. No. Since you have initialized the weights with 1, all the neurons will try
to do the same thing as they will never converge.
Q.23 Suppose that you are training your Artificial Neural Network. After you
have created your model, you evaluate it. However, during model evaluation, you
find out that the training loss/validation loss remains constant. What could be the
reason behind this constant figure loss between training and validation test?
Ans. There are two important reasons that would contribute towards the
training/validation loss stagnation. Firstly, the architecture of the model is not
properly defined. Secondly, the input data has noisy characteristics.
Learn everything about Machine Learning and its Algorithms
Q.24 You have a data science project assignment where you have to deal with
1000 columns and around 1 million rows. The objective of the problem is to
carry out classification. You are required to reduce the dimensions of this data in
order to reduce the model computation time. Furthermore, your machine suffers
from memory constraints. What will you do in this situation?
Ans. Considering memory constraints, developing a machine learning model
would prove to be a laborious task. However, one can carry this out with the
following steps:
 Since we are low on our RAM, we can preserve the memory by closing the
other miscellaneous applications that we do not require.
 We will then perform sampling on our data randomly. This sample will be
a much smaller version of the bigger dataset.
 We will then reduce the dimensionality by removing the correlated
variables. Furthermore, using PCA, we will select those features that can
explain maximum variance in our data.
 We will further create a linear model using stochastic gradient descent.
 Using domain knowledge, we will further drop the predictor variables that
do not have much effect on the response variable. This will further lead to
a reduction in the number of dimensions.
Q.25 Your company has assigned you a new project that involves assisting a food
delivery company to prevent losses from occurring. The main reason behind the
loss is that the food delivery team is not able to deliver food to their customers in
the stipulated time-frame. As a part of their policy, they are then required to
deliver food without any charge. This is resulting in losses on the company’s
part. How can you fix this problem using machine learning algorithm?
Ans. Considering that this question does not have any pattern or required
data, it does not qualify for a machine learning problem. It is clearly a route
optimization problem that will require a different set of algorithms.
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Q.26 Suppose that you have been assigned the task of analyzing text data that is
obtained from the news sentences that are structured. Now, you have to detect
noun phrases, verb phrases as well as perform subject and object detection.
Which grammar-based text parsing technique would you use in this scenario?
Ans. In this scenario, we will make use of Dependency and Constituent
Parsing Extraction techniques to retrieve relations from the textual data.
Q.27 You are working on a Data Science problem in which you have spent a
considerable amount of time in data preprocessing and analysis. You are now
required to implement a machine learning model that would provide you with a
high accuracy. Knowing that boosting algorithms are coveted by data scientists
for their high accuracy, you decide to develop five Gradient Boosting Models.
However, the models do not surpass even the standard benchmark score. You
then create an ensemble of these five models but you do not succeed. Where
exactly did you go wrong?
Ans. Ensemble Learning involves the notion of combining weak learners to
form strong learners. The underlying ensemble models only provide accurate
results when they are uncorrelated. If the ensemble models in the scenario
above do not yield an accurate output then we conclude that the models are
correlated.
Q.28 Suppose that you are working on neural networks where you have to utilise
an activation function in its hidden layers. The output that we obtain is -0.0002.
What type of activation could have been used in order to obtain such type of an
output?
Ans. Since, the output obtained is -0.0002 which is between -1 and 1, the
activation function which has been used in the hidden layer is tanh.
Wait! It is time to revise your neural network concepts.
Q.29 Assume that you are working at DataFlair and you have been assigned the
task of developing a machine learning algorithm that predicts the number of
views an article attracts. During the process of analysis, you include important
features such as author name, number of articles written by the author in the
past etc. What would be the ideal evaluation metric that you would use in this
scenario?
Ans. Number of views that an article attracts on the website is a continuous
target variable which is a part of the regression problem. Therefore, we will
make use of mean squared error as our primary evaluation metric.
Q.30 Assume that you are working with categorical features wherein you do not
know about the distribution of the categorical variable present in the validation
set. Now, you wish to apply one hot encoding on the categorical features. What
are the various challenges that you can encounter once you have applied one hot
encoding on the categorical variable belonging to the train set?
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Ans. Applying One Hot Encoding to encode the categories present in the test
set but not in the train set, will not involve all the categories of the categorical
variable present in the dataset. Secondly, there could be a possible mismatch
between the frequency distribution of the categories present in the training set
and the validation set.
Q.31 Suppose that you have to work with the data present on social media. After
you have retrieved the data have to develop a model that suggests the hashtags to
the user. How will you carry this out?
Ans. We can carry out Topic Modeling to extract significant words present in
the corpus. To capture the top n-gram words and their combinations. And, for
learning repeating contexts in the sentence we train a word2vec model.

Machine Learning and Statistics Data

Science Interview Questions and
Answers
The next important part of our data science interview questions and answers
is mathematics, ML and Statistics. Without having the knowledge of these 3
you cannot become a data scientist. For most of the candidates, statistics
prove as a tough part. So, this is something that can help you to score well in
your data science interview. My tip is to thoroughly learn all the formulas and
definitions related to it.

Q.32 Can you name the type of biases that occur in machine learning?
Ans. There are four main types of biases that occur while building machine
learning algorithms –
 Sample Bias
 Prejudice Bias
 Measurement Bias
 Algorithm Bias
Q.33 How is skewness different from kurtosis?
Ans. In data science, the general meaning of skewness is basically to determine
the imbalance. In statistics, skewness is a measure of asymmetry in the
distribution of data. Ideally, data is normally distributed, meaning that both
the left and right tails are equidistant from the center of the distribution. In
this case, the skewness is 0. However, a distribution exhibits negative
skewness if the left tail is longer than the right one. And, the distribution
exhibits positive skewness if the right tail is longer than the left one.
In case of kurtosis, we measure the pointedness of the peak of distribution.
The ideal kurtosis or the kurtosis of a normal distribution is 3. If the kurtosis
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

of the tail data exceeds 3, then we say that the distributions possess heavy
tails. And, if the kurtosis is less than 3, we say that the distributions have thin
tails.

Q.34 In a univariate linear least squares regression, what is the relationship

between the correlation coefficient and coefficient of determination?
Ans. The relationship between the correlation coefficient and coefficient of
determination in a univariate linear least squares regression is that the latter
is a result of the square of the former. R squared tells us about the coefficient
of determination and it provides a magnitude of variability of the dependent
variable through the independent one.
If there is any concept in Machine learning that you have missed, DataFlair came
with the complete Machine Learning Tutorial Library. Save the page and learn
everything for free at any time.
Q.35 What is z-score?
Ans. Z-score, also known as the standard score is the number of standard
deviations that the data-point is from the mean. It measures how many
standard deviations below or above the population mean is. Z-score ranges
from -3 and goes up till +3 standard deviations.

Q.36 What is the formula of Logistic Regression?

Ans. The formula of Logistic Regression is:

Where P represents the probability, e is the base of natural logarithms and a

and b are the parameters of the logistic regression model.
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Do you know – There is no single Data Science Interview where the question from
logistic regression is not asked. What are you waiting for? Start learning logistic
regression with the best ever guide.
Q.37 For tuning hyperparameters of your machine learning model, what will be
the ideal seed?
Ans. There is no fixed value for the seed and no ideal value. The seed is
initialized randomly in order to tune the hyperparameters of the machine
learning model.
Q.38 Explain the difference between Eigenvalue and Eigenvectors.
Ans. While the eigenvalues are the values that are associated with the degree
of linear transformation, eigenvectors of a non-singular matrix are associated
with its linear transformations that are calculated with correlation or
covariance matrix functions.
Q.39 Is it true that Pearson captures the monotonic behavior of the relation
between the two variables whereas Spearman captures how linearly dependent
the two variables are?
Ans. No. It is actually the opposite. Pearson evaluates the linear relationship
between the two variables whereas Spearman evaluates the monotonic
behavior that the two variables share in a relationship.
Q.40 How is standard deviation affected by the outliers?
In the formula for standard deviation –

The variation in the input value of x, that is, a variation in its value between
high and low would adversely affect the standard deviation and its value would
be farther away from the mean. Therefore, we conclude that outliers will have
an effect on the standard deviation.

Q.41 How will you create a decision tree?

Ans.
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

 If both positive and negative examples are present, we select the attribute
for splitting them.
 If examples are positive, answer yes. Otherwise, answer no.
 When there are no observed examples then we select a default based on
majority classification at the parent.
 If no attributes are remaining, then both the positive and negative
examples are present. This means that there are no sufficient features for
classification or an error is present in the examples.
Master the concept of decision trees and answer all the Data Science Interview
Questions related to it confidently.
Q.42 What is regularization? How is it useful?
Ans. Regularizations are the techniques for reducing the error by fitting a
function on a training set in an appropriate manner to avoid overfitting.
While training the model, there is a high chance of the model learning noise or
the data-points that do not represent any property of your true data. This can
lead to overfitting. Therefore, in order to minimize this form of error, we use
regularization in our machine learning models.

Q.43 Given a linear equation: 2x + 8 = y for the following data-points:

X Y
5 18
6 20
7 22
8 24
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

9 26
What will be the corresponding Mean Absolute Error?
Ans. In order to calculate the mean value error, we first calculate the value of y
as per the given linear equation. Then we calculate the absolute error with
respect to the output value of y. In the end, we find the average of the errors
which is our Mean Absolute Error.
X Y 2x + 8 = y Absolute Error
5 10 18 8
6 21 20 1
7 26 22 4
8 19 24 5
9 30 26 4
Mean Error 4.4
General Data Science Interview Questions and
Answers
Q.44 How is conditional random field different from hidden markov models?
Ans. Conditional Random Fields (CRMs) are discriminative in nature whereas
Hidden Markov Models (HMMs) are generative models.
Q.45 What does the cost parameter in SVM stand for?
Ans. Cost (C) Parameter in SVM decides how well the data should with the
model. Cost Parameter is used for adjusting the hardness or softness of your
large margin classification. With low cost, we make use of a smooth decision
surface whereas to classify more points we make use of the higher cost.
Q.46 Why is gradient descent stochastic in nature?
Ans. The term stochastic means random probability. Therefore, in the case of
stochastic gradient descent, the samples are selected at random instead of
taking the whole in a single iteration.
Q.47 How will you subtract means of each row of matrix?
Ans. In order to subtract the means of each row of a matrix, we will use the
mean() function as follows –
X = [Link](5, 10)

Y = X – [Link](axis=1, keepdims=True)

Q.48 What do you mean by the law of large numbers?

Ans. According to the law of large numbers, the frequency of occurrence of
events that possess the same likelihood are evened out after they undergo a
significant number of trials.
Besant Technologies Contact Us—Karthick Raja -- +91 75500 15337 Marathahalli Branch

Q.49 Explain L1 and L2 Regularization

Both L1 and L2 regularizations are used to avoid overfitting in the model. L1
regularization or Lasso and L2 regularization or Ridge Regularization remove
features from our model. L1 regularization, however, is more tolerant to
outliers. Therefore L1 regularization is much better at handling noisy data.

Q.50 What do the Alpha and Beta Hyperparameter stand for in the Latent
Dirichlet Allocation Model for text classification?
Ans. In the Latent Dirichlet Model for text classification, Alpha represents the
number of topics within the document and Beta stands for the number of
terms occurring within the topic.
Q.51 Is it true that the LogLoss evaluation metric can possess negative values?
Ans. No. Log Loss evaluation metric cannot possess negative values.
Q.52 What is the formula of Stochastic Gradient Descent?
The formula for Stochastic Gradient Descent is as follows:

Q.53 Explain TF/IDF Vectorization

TF/IDF stands for Term Frequency/Inverse Document Frequency. It is used
for information retrieval and mining. It is used as a weighing factor to find the
importance of word to a document. This importance is proportional to, and
increases with the number of times a word occurs in the document but is offset
by the frequency of the word in a corpus.

Q.54 What is Softmax Function? What is the formula of Softmax Normalization?

Ans. Softmax Function is used for normalizing the input into a probability
distribution over the output classes. Following is the formula for the Softmax
Normalization: