Predictive Modeling (19 questions)

1. (Given a Dataset) Analyze this dataset and give me a model that can predict this
response variable.

• Problem Determination -> Data Cleaning -> Feature Engineering -> Modeling
• Benchmark Models
o Linear Regression (Ridge or Lasso) for regression
o Logistic Regression for Classification
• Advanced Models
o Random Forest, Boosting Trees, and so on
 Scikit-Learn, XGBoost, LightGBM, CatBoost
• Determine if the problem is classification or regression
• Plot and visualize the data.
• Start by fitting a simple model (multivariate regression, logistic regression), do
some feature engineering accordingly, and then try some complicated models.
Always split the dataset into train, validation, test dataset and use cross validation
to check their performance.
• Favor simple models that run quickly and you can easily explain.
• Mention cross validation as a means to evaluate the model.

2. What could be some issues if the distribution of the test data is significantly
different than the distribution of the training data?

• The model that has high training accuracy might have low test accuracy. Without
further knowledge, it is hard to know which dataset represents the population
data and thus the generalizability of the algorithm is hard to measure. This
should be mitigated by repeated splitting of train vs. test dataset (as in cross
validation).
• When there is a change in data distribution, this is called the dataset shift. If the
train and test data has a different distribution, then the classifier would likely
overfit to the train data.
• This issue can be overcome by using a more general learning method.
• This can occur when:
o P(y|x) are the same but P(x) are different. (covariate shift)
o P(y|x) are different. (concept shift)
• The causes can be:
 o Training samples are obtained in a biased way. (sample selection bias)
o Train is different from test because of temporal, spatial changes. (non-
stationary environments)
• Solution to covariate shift
o importance weighted cv

3. What are some ways I can make my model more robust to outliers?

• We can have regularization such as L1 or L2 to reduce variance (increase bias).

• Changes to the algorithm:
o Use tree-based methods instead of regression methods as they are more
resistant to outliers. For statistical tests, use non parametric tests instead of
parametric ones.
o Use robust error metrics such as MAE or Huber Loss instead of MSE.
• Changes to the data:
o Winsorizing the data
o Transforming the data (e.g. log)
o Remove them only if you’re certain they’re anomalies and not worth
predicting

4. What are some differences you would expect in a model that minimizes squared
error, versus a model that minimizes absolute error? In which cases would each
error metric be appropriate?

• MSE is more strict to having outliers. MAE is more robust in that sense, but is
harder to fit the model for because it cannot be numerically optimized. So when
there are less variability in the model and the model is computationally easy to
fit, we should use MAE, and if that’s not the case, we should use MSE.
• MSE: easier to compute the gradient, MAE: linear programming needed to
compute the gradient
• MAE more robust to outliers. If the consequences of large errors are great, use
MSE
• MSE corresponds to maximizing likelihood of Gaussian random variables

5. What error metric would you use to evaluate how good a binary classifier is?
What if the classes are imbalanced? What if there are more than 2 groups?

• Accuracy: proportion of instances you predict correctly.

 o Pros: intuitive, easy to explain
o Cons: works poorly when the class labels are imbalanced and the signal
from the data is weak
• ROC curve and AUC: plot false-positive-rate (fpr) on the x axis and true-positive-
rate (tpr) on the y axis for different threshold. Given a random positive instance
and a random negative instance, the AUC is the probability that you can identify
who's who.
o Pros: Works well when testing the ability of distinguishing the two classes.
o Cons: can’t interpret predictions as probabilities (because AUC is
determined by rankings), so can’t explain the uncertainty of the model,
and it doesn't work for multi-class case.
• logloss/deviance/cross entropy:
o Pros: error metric based on probabilities
o Cons: very sensitive to false positives, negatives
• When there are more than 2 groups, we can have k binary classifications and add
them up for logloss. Some metrics like AUC is only applicable in the binary case.

6. What are various ways to predict a binary response variable? Can you compare
two of them and tell me when one would be more appropriate? What’s the
difference between these? (SVM, Logistic Regression, Naive Bayes, Decision Tree,
etc.)

• Things to look at: N, P, linearly separable, features independent, likely to overfit,

speed, performance, memory usage and so on.
• Logistic Regression
o features roughly linear, problem roughly linearly separable
o robust to noise, use l1,l2 regularization for model selection, avoid
overfitting
o the output come as probabilities
o efficient and the computation can be distributed
o can be used as a baseline for other algorithms
o (-) can hardly handle categorical features
• SVM
o with a nonlinear kernel, can deal with problems that are not linearly
separable
o (-) slow to train, for most industry scale applications, not really efficient
• Naive Bayes
 o computationally efficient when P is large by alleviating the curse of
dimensionality
o works surprisingly well for some cases even if the condition doesn’t hold
o with word frequencies as features, the independence assumption can be
seen reasonable. So the algorithm can be used in text categorization
o (-) conditional independence of every other feature should be met
• Tree Ensembles
o good for large N and large P, can deal with categorical features very well
o non parametric, so no need to worry about outliers
o GBT’s work better but the parameters are harder to tune
o RF works out of the box, but usually performs worse than GBT
• Deep Learning
o works well for some classification tasks (e.g. image)
o used to squeeze something out of the problem

7. What is regularization and where might it be helpful? What is an example of

using regularization in a model?

• Regularization is useful for reducing variance in the model, meaning avoiding

overfitting.
• For example, we can use L1 regularization in Lasso regression to penalize large
coefficients and automatically select features, or we can also use L2 regularization
for Ridge regression to penalize the feature coefficients.

8. Why might it be preferable to include fewer predictors over many?

• When we add irrelevant features, it increases model's tendency to overfit because

those features introduce more noise. When two variables are correlated, they
might be harder to interpret in case of regression, etc.
• curse of dimensionality
• adding random noise makes the model more complicated but useless
• computational cost
• Ask someone for more details.

9. Given training data on tweets and their retweets, how would you predict the
number of retweets of a given tweet after 7 days after only observing 2 days
worth of data?
 • Build a time series model with the training data with a seven day cycle and then
use that for a new data with only 2 days data.
• Ask someone for more details.
• Build a regression function to estimate the number of retweets as a function of
time t
• to determine if one regression function can be built, see if there are clusters in
terms of the trends in the number of retweets
• if not, we have to add features to the regression function
• features + # of retweets on the first and the second day -> predict the seventh
day
• https://en.wikipedia.org/wiki/Dynamic_time_warping

10. How could you collect and analyze data to use social media to predict the
weather?

• We can collect social media data using twitter, Facebook, instagram API’s.
• Then, for example, for twitter, we can construct features from each tweet, e.g. the
tweeted date, number of favorites, retweets, and of course, the features created
from the tweeted content itself.
• Then use a multivariate time series model to predict the weather.
• Ask someone for more details.

11. How would you construct a feed to show relevant content for a site that
involves user interactions with items?

• We can do so using building a recommendation engine.

• The easiest we can do is to show contents that are popular other users, which is
still a valid strategy if for example the contents are news articles.
• To be more accurate, we can build a content based filtering or collaborative
filtering. If there’s enough user usage data, we can try collaborative filtering and
recommend contents other similar users have consumed. If there isn’t, we can
recommend similar items based on vectorization of items (content based
filtering).

12. How would you design the people you may know feature on LinkedIn or
Facebook?

• Find strong unconnected people in weighted connection graph

 o Define similarity as how strong the two people are connected
o Given a certain feature, we can calculate the similarity based on
 friend connections (neighbors)
 Check-in’s people being at the same location all the time.
 same college, workplace
 Have randomly dropped graphs test the performance of the
algorithm
• Ref. News Feed Optimization
o Affinity score: how close the content creator and the users are
o Weight: weight for the edge type (comment, like, tag, etc.). Emphasis on
features the company wants to promote
o Time decay: the older the less important

13. How would you predict who someone may want to send a Snapchat or Gmail
to?

• for each user, assign a score of how likely someone would send an email to
• the rest is feature engineering:
o number of past emails, how many responses, the last time they exchanged
an email, whether the last email ends with a question mark, features about
the other users, etc.
• Ask someone for more details.
• People who someone sent emails the most in the past, conditioning on time
decay.

14. How would you suggest to a franchise where to open a new store?

• build a master dataset with local demographic information available for each
location.
o local income levels, proximity to traffic, weather, population density,
proximity to other businesses
o a reference dataset on local, regional, and national macroeconomic
conditions (e.g. unemployment, inflation, prime interest rate, etc.)
o any data on the local franchise owner-operators, to the degree the
manager
• identify a set of KPIs acceptable to the management that had requested the
analysis concerning the most desirable factors surrounding a franchise
o quarterly operating profit, ROI, EVA, pay-down rate, etc.
 • run econometric models to understand the relative significance of each variable
• run machine learning algorithms to predict the performance of each location
candidate

15. In a search engine, given partial data on what the user has typed, how would
you predict the user’s eventual search query?

• Based on the past frequencies of words shown up given a sequence of words, we

can construct conditional probabilities of the set of next sequences of words that
can show up (n-gram). The sequences with highest conditional probabilities can
show up as top candidates.
• To further improve this algorithm,
o we can put more weight on past sequences which showed up more
recently and near your location to account for trends
o show your recent searches given partial data
• Personalize and localize the search
o Use the user's historical search data
o Use the historical data from the local region

16. Given a database of all previous alumni donations to your university, how
would you predict which recent alumni are most likely to donate?

• Based on frequency and amount of donations, graduation year, major, etc,

construct a supervised regression (or binary classification) algorithm.

17. You’re Uber and you want to design a heatmap to recommend to drivers where
to wait for a passenger. How would you approach this?

• Based on the past pickup location of passengers around the same time of the
day, day of the week (month, year), construct
• Ask someone for more details.
• Based on the number of past pickups
o account for periodicity (seasonal, monthly, weekly, daily, hourly)
o special events (concerts, festivals, etc.) from tweets

18. How would you build a model to predict a March Madness bracket?

• One vector each for team A and B. Take the difference of the two vectors and use
that as an input to predict the probability that team A would win by training the
 model. Train the models using past tournament data and make a prediction for
the new tournament by running the trained model for each round of the
tournament
• Some extensions:
o Experiment with different ways of consolidating the 2 team vectors into
one (e.g concantenating, averaging, etc)
o Consider using a RNN type model that looks at time series data.

19. You want to run a regression to predict the probability of a flight delay, but
there are flights with delays of up to 12 hours that are really messing up your
model. How can you address this?

• This is equivalent to making the model more robust to outliers.

• See Q3.
Programming (14 questions)

1. Write a function to calculate all possible assignment vectors of 2n users, where

n users are assigned to group 0 (control), and n users are assigned to group 1
(treatment).

• Recursive programming (sol in code)

def n_choose_k(n, k):

""" function to choose k from n """
if k == 1:
ans = []
for i in range(n):
tmp = [0] * n
tmp[i] = 1
ans.append(tmp)
return ans

if k == n:
return [[1] * n]

ans = []
space = n - k + 1
for i in range(space):
assignment = [0] * (i + 1)
assignment[i] = 1
for c in n_choose_k(n - i - 1, k - 1):
ans.append(assignment + c)
return ans

# test: choose 2 from 4

print(n_choose_k(4, 2))

2. Given a list of tweets, determine the top 10 most used hashtags.

• Store all the hashtags in a dictionary and use priority queue to solve the top-k
problem
• An extension will be top-k problem using Hadoop/MapReduce

3. Program an algorithm to find the best approximate solution to the knapsack

problem in a given time.

• https://en.wikipedia.org/wiki/Knapsack_problem
• Greedy solution (add the best v/w as much as possible and move on to the next)
• Dynamic programming
4. Program an algorithm to find the best approximate solution to the traveling
salesman problem in a given time.

• https://en.wikipedia.org/wiki/Travelling_salesman_problem
• Greedy
• Dynamic programming

5. You have a stream of data coming in of size n, but you don’t know what n is
ahead of time. Write an algorithm that will take a random sample of k elements.
Can you write one that takes O(k) space?

• Reservoir sampling

6. Write an algorithm that can calculate the square root of a number.

• Binary search or Newton's method

7. Given a list of numbers, can you return the outliers?

• Sort then select the highest and the lowest 2.5%

• Visualization can helps a lot

8. When can parallelism make your algorithms run faster? When could it make
your algorithms run slower?

• Ask someone for more details.

• compute in parallel when communication cost < computation cost
o ensemble trees
o minibatch
o cross validation
o forward propagation
o minibatch
o not suitable for online learning

9. What are the different types of joins? What are the differences between them?

• Inner Join, Left Join, Right Join, Outer Join, Self Join
10. Why might a join on a subquery be slow? How might you speed it up?

• Change the subquery to a join.

• Stack Overflow Answers

11. Describe the difference between primary keys and foreign keys in a SQL
database.

• Primary keys are columns whose value combinations must be unique in a specific
table so that each row can be referenced uniquely.
• Foreign keys are columns that references columns (often primary keys) in other
tables.

12. Given a COURSES table with columns course_id and course_name,

a FACULTY table with columns faculty_id and faculty_name, and
a COURSE_FACULTY table with columns faculty_id and course_id, how would you
return a list of faculty who teach a course given the name of a course?

SELECT f.faculty_name
FROM COURSES c
JOIN COURSE_FACULTY cf
ON c.course_id = cf.course_id
JOIN FACULTY
ON f.faculty_id = cf.faculty_id
WHERE c.course_name = xxx;

13. Given a IMPRESSIONS table with ad_id, click (an indicator that the ad was
clicked), and date, write a SQL query that will tell me the click-through-rate of
each ad by month.

SELECT ad_id, MONTH(date), AVG(click)

FROM IMPRESSIONS
GROUP BY ad_id, MONTH(date);

14. Write a query that returns the name of each department and a count of the
number of employees in each:

• EMPLOYEES containing: Emp_ID (Primary key) and Emp_Name

• EMPLOYEE_DEPT containing: Emp_ID (Foreign key) and Dept_ID (Foreign key)
• DEPTS containing: Dept_ID (Primary key) and Dept_Name

SELECT d.Dept_Name, COUNT(*)

FROM DEPTS d
LEFT JOIN EMPLOYEE_DEPT ed
ON d.Dept_ID = ed.Dept_ID
GROUP BY d.Dept_Name;
Probability (19 questions)

1. Bobo the amoeba has a 25%, 25%, and 50% chance of producing 0, 1, or 2 o
spring, respectively. Each of Bobo’s descendants also have the same probabilities.
What is the probability that Bobo’s lineage dies out?

• p=1/4+1/4p+1/2p^2 => p=1/2

2. In any 15-minute interval, there is a 20% probability that you will see at least
one shooting star. What is the probability that you see at least one shooting star in
the period of an hour?

• 1-(0.8)^4 = 0.5904
• Or, we can use Poisson processes

3. How can you generate a random number between 1 - 7 with only a die?

• Quora Answer

4. How can you get a fair coin toss if someone hands you a coin that is weighted to
come up heads more often than tails?

• Flip twice:
o HT --> H
o TH --> T
• If HH or TT, repeat.

5. You have an 50-50 mixture of two normal distributions with the same standard
deviation. How far apart do the means need to be in order for this distribution to
be bimodal?

• more than two standard deviations

6. Given draws from a normal distribution with known parameters, how can you
simulate draws from a uniform distribution?

• Plug in the value to the CDF of the same random variable

7. A certain couple tells you that they have two children, at least one of which is a
girl. What is the probability that they have two girls?

• gg, gb, bg --> 1/3

8. You have a group of couples that decide to have children until they have their
first girl, after which they stop having children. What is the expected gender ratio
of the children that are born? What is the expected number of children each
couple will have?

• Geometric distribution with p = 0.5

• gender ratio is 1:1. Expected number of children is 2.
• let X be the number of children until getting a female (happens with prob 1/2).
this follows a geometric distribution with probability 1/2

9. How many ways can you split 12 people into 3 teams of 4?

• the outcome follows a multinomial distribution with n=12 and k=3. but the
classes are indistinguishable
• (12, 8) * (8, 4) * (4, 4) / (3, 3)
• 12! / (4!)^3 / 3!

10. Your hash function assigns each object to a number between 1:10, each with
equal probability. With 10 objects, what is the probability of a hash collision?
What is the expected number of hash collisions? What is the expected number of
hashes that are unused.

• the probability of a hash collision: 1-(10!/10^10)

• the expected number of hash collisions: 10(1 - (1-1/10)^10)
o Quora Reference
• the expected number of hashes that are unused: 10*(9/10)^10

11. You call 2 UberX’s and 3 Lyfts. If the time that each takes to reach you is IID,
what is the probability that all the Lyfts arrive first? What is the probability that all
the UberX’s arrive first?

• Lyfts arrive first: 2! * 3! / 5!

• Ubers arrive first: same
12. I write a program should print out all the numbers from 1 to 300, but prints
out Fizz instead if the number is divisible by 3, Buzz instead if the number is
divisible by 5, and FizzBuzz if the number is divisible by 3 and 5. What is the total
number of numbers that is either Fizzed, Buzzed, or FizzBuzzed?

• 100+60-20=140

13. On a dating site, users can select 5 out of 24 adjectives to describe themselves.
A match is declared between two users if they match on at least 4 adjectives. If
Alice and Bob randomly pick adjectives, what is the probability that they form a
match?

• 24C5*(1+5(24-5))/24C5*24C5 = 4/1771

14. A lazy high school senior types up application and envelopes to n different
colleges, but puts the applications randomly into the envelopes. What is the
expected number of applications that went to the right college?

• 1

15. Let’s say you have a very tall father. On average, what would you expect the
height of his son to be? Taller, equal, or shorter? What if you had a very short
father?

• Shorter. Regression to the mean

16. What’s the expected number of coin flips until you get two heads in a row?
What’s the expected number of coin flips until you get two tails in a row?

• x = 0.25 * 2 + 0.25 * (x + 2) + 0.5 * (x + 1) --> x = 6

• Quora Reference

17. Let’s say we play a game where I keep flipping a coin until I get heads. If the
first time I get heads is on the nth coin, then I pay you 2n-1 dollars. How much
would you pay me to play this game?

• less than $3
• Quora reference
18. You have two coins, one of which is fair and comes up heads with a probability
1/2, and the other which is biased and comes up heads with probability 3/4. You
randomly pick coin and flip it twice, and get heads both times. What is the
probability that you picked the fair coin?

• 4/13
• Bayesian method

19. You have a 0.1% chance of picking up a coin with both heads, and a 99.9%
chance that you pick up a fair coin. You flip your coin and it comes up heads 10
times. What’s the chance that you picked up the fair coin, given the information
that you observed?

• Bayesian method

20. What is a P-Value ?

• https://en.wikipedia.org/wiki/P-value
Statistical Inference (15 questions)

1. In an A/B test, how can you check if assignment to the various buckets was truly
random?

• Plot the distributions of multiple features for both A and B and make sure that
they have the same shape. More rigorously, we can conduct a permutation test to
see if the distributions are the same.
• MANOVA to compare different means

2. What might be the benefits of running an A/A test, where you have two buckets
who are exposed to the exact same product?

• Verify the sampling algorithm is random.

3. What would be the hazards of letting users sneak a peek at the other bucket in
an A/B test?

• The user might not act the same suppose had they not seen the other bucket.
You are essentially adding additional variables of whether the user peeked the
other bucket, which are not random across groups.

4. What would be some issues if blogs decide to cover one of your experimental
groups?

• Same as the previous question. The above problem can happen in larger scale.

5. How would you conduct an A/B test on an opt-in feature?

• Ask someone for more details.

6. How would you run an A/B test for many variants, say 20 or more?

• one control, 20 treatment, if the sample size for each group is big enough.
• Ways to attempt to correct for this include changing your confidence level (e.g.
Bonferroni Correction) or doing family-wide tests before you dive in to the
individual metrics (e.g. Fisher's Protected LSD).

7. How would you run an A/B test if the observations are extremely right-skewed?
 • lower the variability by modifying the KPI
• cap values
• percentile metrics
• log transform
• https://www.quora.com/How-would-you-run-an-A-B-test-if-the-observations-
are-extremely-right-skewed

8. I have two different experiments that both change the sign-up button to my
website. I want to test them at the same time. What kinds of things should I keep
in mind?

• exclusive -> ok

9. What is a p-value? What is the difference between type-1 and type-2 error?

• en.wikipedia.org/wiki/P-value
• type-1 error: rejecting Ho when Ho is true
• type-2 error: not rejecting Ho when Ha is true

10. You are AirBnB and you want to test the hypothesis that a greater number of
photographs increases the chances that a buyer selects the listing. How would you
test this hypothesis?

• For randomly selected listings with more than 1 pictures, hide 1 random picture
for group A, and show all for group B. Compare the booking rate for the two
groups.
• Ask someone for more details.

11. How would you design an experiment to determine the impact of latency on
user engagement?

• The best way I know to quantify the impact of performance is to isolate just that
factor using a slowdown experiment, i.e., add a delay in an A/B test.

12. What is maximum likelihood estimation? Could there be any case where it
doesn’t exist?
 • A method for parameter optimization (fitting a model). We choose parameters so
as to maximize the likelihood function (how likely the outcome would happen
given the current data and our model).
• maximum likelihood estimation (MLE) is a method
of estimating the parameters of a statistical model given observations, by finding
the parameter values that maximize the likelihood of making the observations
given the parameters. MLE can be seen as a special case of the maximum a
posteriori estimation (MAP) that assumes a uniform prior distribution of the
parameters, or as a variant of the MAP that ignores the prior and which therefore
is unregularized.
• for gaussian mixtures, non parametric models, it doesn’t exist

13. What’s the difference between a MAP, MOM, MLE estimator? In which cases
would you want to use each?

• MAP estimates the posterior distribution given the prior distribution and data
which maximizes the likelihood function. MLE is a special case of MAP where the
prior is uninformative uniform distribution.
• MOM sets moment values and solves for the parameters. MOM is not used much
anymore because maximum likelihood estimators have higher probability of
being close to the quantities to be estimated and are more often unbiased.

14. What is a confidence interval and how do you interpret it?

• For example, 95% confidence interval is an interval that when constructed for a
set of samples each sampled in the same way, the constructed intervals include
the true mean 95% of the time.
• if confidence intervals are constructed using a given confidence level in an infinite
number of independent experiments, the proportion of those intervals that
contain the true value of the parameter will match the confidence level.

15. What is unbiasedness as a property of an estimator? Is this always a desirable

property when performing inference? What about in data analysis or predictive
modeling?

• Unbiasedness means that the expectation of the estimator is equal to the

population value we are estimating. This is desirable in inference because the
goal is to explain the dataset as accurately as possible. However, this is not
always desirable for data analysis or predictive modeling as there is the bias
variance tradeoff. We sometimes want to prioritize the generalizability and avoid
overfitting by reducing variance and thus increasing bias.
Data Analysis (27 questions)

1. (Given a Dataset) Analyze this dataset and tell me what you can learn from it.

• Typical data cleaning and visualization

2. What is R2? What are some other metrics that could be better than R2 and why?

• goodness of fit measure. variance explained by the regression / total variance

• the more predictors you add, the higher R^2 becomes.
o hence use adjusted R^2 which adjusts for the degrees of freedom
o or train error metrics

3. What is the curse of dimensionality?

• High dimensionality makes clustering hard, because having lots of dimensions

means that everything is "far away" from each other.
• For example, to cover a fraction of the volume of the data we need to capture a
very wide range for each variable as the number of variables increases
• All samples are close to the edge of the sample. And this is a bad news because
prediction is much more difficult near the edges of the training sample.
• The sampling density decreases exponentially as p increases and hence the data
becomes much more sparse without significantly more data.
• We should conduct PCA to reduce dimensionality

4. Is more data always better?

• Statistically
o It depends on the quality of your data, for example, if your data is biased,
just getting more data won’t help.
o It depends on your model. If your model suffers from high bias, getting
more data won’t improve your test results beyond a point. You’d need to
add more features, etc.
• Practically
o More data usually benefit the models
o Also there’s a tradeoff between having more data and the additional
storage, computational power, memory it requires. Hence, always think
about the cost of having more data.
5. What are advantages of plotting your data before performing analysis?

• Data sets have errors. You won't find them all but you might find some. That 212
year old man. That 9 foot tall woman.
• Variables can have skewness, outliers, etc. Then the arithmetic mean might not be
useful, which means the standard deviation isn't useful.
• Variables can be multimodal! If a variable is multimodal then anything based on
its mean or median is going to be suspect.

6. How can you make sure that you don’t analyze something that ends up
meaningless?

• Proper exploratory data analysis.

• In every data analysis task, there's the exploratory phase where you're just
graphing things, testing things on small sets of the data, summarizing simple
statistics, and getting rough ideas of what hypotheses you might want to pursue
further.
• Then there's the exploratory phase, where you look deeply into a set of
hypotheses.
• The exploratory phase will generate lots of possible hypotheses, and the
exploratory phase will let you really understand a few of them. Balance the two
and you'll prevent yourself from wasting time on many things that end up
meaningless, although not all.

7. What is the role of trial and error in data analysis? What is the the role of
making a hypothesis before diving in?

• data analysis is a repetition of setting up a new hypothesis and trying to refute

the null hypothesis.
• The scientific method is eminently inductive: we elaborate a hypothesis, test it
and refute it or not. As a result, we come up with new hypotheses which are in
turn tested and so on. This is an iterative process, as science always is.

8. How can you determine which features are the most important in your model?

• Linear regression can use p-value

• run the features though a Gradient Boosting Machine or Random Forest to
generate plots of relative importance and information gain for each feature in the
ensembles.
 • Look at the variables added in forward variable selection

9. How do you deal with some of your predictors being missing?

• Remove rows with missing values - This works well if

o the values are missing randomly (see Vinay Prabhu's answer for more
details on this)
o if you don't lose too much of the dataset after doing so.
• Build another predictive model to predict the missing values
o This could be a whole project in itself, so simple techniques are usually
used here.
• Use a model that can incorporate missing data
o Like a random forest, or any tree-based method.

10. You have several variables that are positively correlated with your response,
and you think combining all of the variables could give you a good prediction of
your response. However, you see that in the multiple linear regression, one of the
weights on the predictors is negative. What could be the issue?

• Multicollinearity refers to a situation in which two or more explanatory variables

in a multiple regression model are highly linearly related.
• Leave the model as is, despite multicollinearity. The presence of multicollinearity
doesn't affect the efficiency of extrapolating the fitted model to new data
provided that the predictor variables follow the same pattern of multicollinearity
in the new data as in the data on which the regression model is based.
• principal component regression

11. Let’s say you’re given an unfeasible amount of predictors in a predictive

modeling task. What are some ways to make the prediction more feasible?

• PCA

12. Now you have a feasible amount of predictors, but you’re fairly sure that you
don’t need all of them. How would you perform feature selection on the dataset?

• ridge / lasso / elastic net regression

• Univariate Feature Selection where a statistical test is applied to each feature
individually. You retain only the best features according to the test outcome
scores
 • Recursive Feature Elimination:
o First, train a model with all the feature and evaluate its performance on
held out data.
o Then drop let say the 10% weakest features (e.g. the feature with least
absolute coefficients in a linear model) and retrain on the remaining
features.
o Iterate until you observe a sharp drop in the predictive accuracy of the
model.

13. Your linear regression didn’t run and communicates that there are an infinite
number of best estimates for the regression coefficients. What could be wrong?

• p > n.
• If some of the explanatory variables are perfectly correlated (positively or
negatively) then the coefficients would not be unique.

14. You run your regression on different subsets of your data, and find that in each
subset, the beta value for a certain variable varies wildly. What could be the issue
here?

• The dataset might be heterogeneous. In which case, it is recommended to cluster

datasets into different subsets wisely, and then draw different models for
different subsets. Or, use models like non parametric models (trees) which can
deal with heterogeneity quite nicely.

15. What is the main idea behind ensemble learning? If I had many different
models that predicted the same response variable, what might I want to do to
incorporate all of the models? Would you expect this to perform better than an
individual model or worse?

• The assumption is that a group of weak learners can be combined to form a

strong learner.
• Hence the combined model is expected to perform better than an individual
model.
• Assumptions:
o average out biases
o reduce variance
• Bagging works because some underlying learning algorithms are unstable:
slightly different inputs leads to very different outputs. If you can take advantage
 of this instability by running multiple instances, it can be shown that the reduced
instability leads to lower error. If you want to understand why, the original
bagging paper( http://www.springerlink.com/) has a section called "why bagging
works"
• Boosting works because of the focus on better defining the "decision edge". By
re-weighting examples near the margin (the positive and negative examples) you
get a reduced error (see http://citeseerx.ist.psu.edu/vie...)
• Use the outputs of your models as inputs to a meta-model.

For example, if you're doing binary classification, you can use all the probability outputs
of your individual models as inputs to a final logistic regression (or any model, really)
that can combine the probability estimates.

One very important point is to make sure that the output of your models are out-of-
sample predictions. This means that the predicted value for any row in your data-frame
should NOT depend on the actual value for that row.

16. Given that you have wifi data in your office, how would you determine which
rooms and areas are underutilized and over-utilized?

• If the data is more used in one room, then that one is over utilized!
• Maybe account for the room capacity and normalize the data.

17. How could you use GPS data from a car to determine the quality of a driver?

• Speed
• Driving paths

18. Given accelerometer, altitude, and fuel usage data from a car, how would you
determine the optimum acceleration pattern to drive over hills?

• Historical data?

19. Given position data of NBA players in a season’s games, how would you
evaluate a basketball player’s defensive ability?

• Evaluate his positions in the court.

20. How would you quantify the influence of a Twitter user?

 • like page rank with each user corresponding to the webpages and linking to the
page equivalent to following.

21. Given location data of golf balls in games, how would construct a model that
can advise golfers where to aim?

• winning probability for different positions

22. You have 100 mathletes and 100 math problems. Each mathlete gets to choose
10 problems to solve. Given data on who got what problem correct, how would
you rank the problems in terms of difficulty?

• One way you could do this is by storing a "skill level" for each user and a
"difficulty level" for each problem. We assume that the probability that a user
solves a problem only depends on the skill of the user and the difficulty of the
problem.* Then we maximize the likelihood of the data to find the hidden skill
and difficulty levels.
• The Rasch model for dichotomous data takes the form:
{\displaystyle \Pr\{X_{ni}=1\}={\frac {\exp({\beta_{n}}-
{\delta_{i}})}{1+\exp({\beta_{n}}-{\delta_{i}})}},}
where is the ability of person and is the difficulty of item}.

23. You have 5000 people that rank 10 sushis in terms of saltiness. How would you
aggregate this data to estimate the true saltiness rank in each sushi?

• Some people would take the mean rank of each sushi. If I wanted something
simple, I would use the median, since ranks are (strictly speaking) ordinal and not
interval, so adding them is a bit risque (but people do it all the time and you
probably won't be far wrong).

24. Given data on congressional bills and which congressional representatives co-
sponsored the bills, how would you determine which other representatives are
most similar to yours in voting behavior? How would you evaluate who is the most
liberal? Most republican? Most bipartisan?

• collaborative filtering. you have your votes and we can calculate the similarity for
each representatives and select the most similar representative
• for liberal and republican parties, find the mean vector and find the
representative closest to the center point
25. How would you come up with an algorithm to detect plagiarism in online
content?

• reduce the text to a more compact form (e.g. fingerprinting, bag of words) then
compare those with other texts by calculating the similarity

26. You have data on all purchases of customers at a grocery store. Describe to me
how you would program an algorithm that would cluster the customers into
groups. How would you determine the appropriate number of clusters to include?

• K-means
• choose a small value of k that still has a low SSE (elbow method)
• Elbow method

27. Let’s say you’re building the recommended music engine at Spotify to
recommend people music based on past listening history. How would you
approach this problem?

• content-based filtering
• collaborative filtering
Product Metrics (15 questions)

1. What would be good metrics of success for an advertising-driven consumer

product? (Buzzfeed, YouTube, Google Search, etc.) A service-driven consumer
product? (Uber, Flickr, Venmo, etc.)

• advertising-driven: Page-views and daily actives, CTR, CPC (cost per click)
o click-ads
o display-ads
• service-driven: number of purchases, conversion rate

2. What would be good metrics of success for a productivity tool? (Evernote,

Asana, Google Docs, etc.) A MOOC? (edX, Coursera, Udacity, etc.)

• Productivity tool: same as premium subscriptions

• MOOC: same as premium subscriptions, completion rate

3. What would be good metrics of success for an e-commerce product? (Etsy,

Groupon, Birchbox, etc.) A subscription product? (Net ix, Birchbox, Hulu, etc.)
Premium subscriptions? (OKCupid, LinkedIn, Spotify, etc.)

• e-commerce: number of purchases, conversion rate, Hourly, daily, weekly,

monthly, quarterly, and annual sales, Cost of goods sold, Inventory levels, Site
traffic, Unique visitors versus returning visitors, Customer service phone call
count, Average resolution time
• subscription
o churn, CoCA, ARPU, MRR, LTV
• premium subscriptions:
o subscription rate

4. What would be good metrics of success for a consumer product that relies
heavily on engagement and interaction? (Snapchat, Pinterest, Facebook, etc.) A
messaging product? (GroupMe, Hangouts, Snapchat, etc.)

• heavily on engagement and interaction: uses AU ratios, email summary by type,

and push notification summary by type, resurrection ratio
• messaging product:
o daily, monthly active users
5. What would be good metrics of success for a product that offered in-app
purchases? (Zynga, Angry Birds, other gaming apps)

• Average Revenue Per Paid User

• Average Revenue Per User

6. A certain metric is violating your expectations by going down or up more than

you expect. How would you try to identify the cause of the change?

• breakdown the KPI’s into what consists them and find where the change is
• then further breakdown that basic KPI by channel, user cluster, etc. and relate
them with any campaigns, changes in user behaviors in that segment

7. Growth for total number of tweets sent has been slow this month. What data
would you look at to determine the cause of the problem?

• Historical data, especially historical data at the same month

• Outer data, such as economic data, political data, data about competitors

8. You’re a restaurant and are approached by Groupon to run a deal. What data
would you ask from them in order to determine whether or not to do the deal?

• for similar restaurants (they should define similarity), average increase in revenue
gain per coupon, average increase in customers per coupon

9. You are tasked with improving the efficiency of a subway system. Where would
you start?

• define efficiency

10. Say you are working on Facebook News Feed. What would be some metrics
that you think are important? How would you make the news each person gets
more relevant?

• rate for each action, duration users stay, CTR for sponsor feed posts
• ref. News Feed Optimization
o Affinity score: how close the content creator and the users are
o Weight: weight for the edge type (comment, like, tag, etc.). Emphasis on
features the company wants to promote
 o Time decay: the older the less important

11. How would you measure the impact that sponsored stories on Facebook News
Feed have on user engagement? How would you determine the optimum balance
between sponsored stories and organic content on a user’s News Feed?

• AB test on different balance ratio and see

12. You are on the data science team at Uber and you are asked to start thinking
about surge pricing. What would be the objectives of such a product and how
would you start looking into this?

• there is a gradual step-function type scaling mechanism until that imbalance of

requests-to-drivers is alleviated and then vice versa as too many drivers come
online enticed by the surge pricing structure.
• I would bet the algorithm is custom tailored and calibrated to each location as
price elasticities almost certainly vary across different cities depending on a huge
multitude of variables: income, distance/sprawl, traffic patterns, car ownership,
etc. With the massive troves of user data that Uber probably has collected, they
most likely have tweaked the algorithms for each city to adjust for these varying
sensitivities to surge pricing. Throw in some machine learning and incredibly rich
data and you've got yourself an incredible, constantly-evolving algorithm.

13. Say that you are Netflix. How would you determine what original series you
should invest in and create?

• Netflix uses data to estimate the potential market size for an original series
before giving it the go-ahead.

14. What kind of services would find churn (metric that tracks how many
customers leave the service) helpful? How would you calculate churn?

• subscription based services

15. Let’s say that you’re are scheduling content for a content provider on
television. How would you determine the best times to schedule content?

• Based on similar product and the corresponding broadcast popularity

1. Explain to me a technical concept related to the role that you’re interviewing
for.

• AB test, PCA, data science, machine learning, neural networks

2. Introduce me to something you’re passionate about.

• Data science

3. How would you explain an A/B test to an engineer with no statistics

background? A linear regression?

• A/B testing, or more broadly, multivariate testing, is the testing of different

elements of a user's experience to determine which variation helps the business
achieve its goal more effectively (i.e. increasing conversions, etc..) This can be
copy on a web site, button colors, different user interfaces, different email subject
lines, calls to action, offers, etc.

4. How would you explain a confidence interval to an engineer with no statistics

background? What does 95% confidence mean?

Suppose we are interested some characteristic of a population; for example, the average
height h of all adult males in the U.S. We can estimate h by drawing a random sample of
adult males in the U.S. and calculating the average height H in the sample. This is called
a point estimate of h. If the sample is large, H will be a good estimate of h, but by itself it
does not tell you how good it is.

A 95% confidence interval is a different kind of estimate. It consists

of two numbers L (lower) and U (upper), which are derived from the sample in some way
without knowledge of the unknown h (or any other unknown parameters). The interval
(L,U) is supposed to contain the unknown h. A procedure for finding (L,U) which does in
fact contain h for 95% of the possible samples is called a 95% confidence interval. If the
interval is short, it gives us a small range of "likely" values for h.

That is the definition. Now, a few comments. Why the strange word "confidence," which is
never used by itself in probability or statistics? Why the scare quotes around the word
"likely" in the previous paragraph?

Confidence intervals are a tool of the frequentist school of statistics, which holds that we
should use the concepts of probability and randomness only to describe the mechanics of
certain kinds of sampling from populations, and not to describe our certainty or degree of
belief. Frequentists aim to use probability in an objective way.

For a frequentist, a statement like "the probability that the average height h of the all males
in the US lies between 70 and 74 inches is 95%" is meaningless: h is just a number we don't
know. It either lies in the interval (70, 74) or it doesn't.

Confidence intervals are a trick frequentists use to make statements resembling the one
above without violating their rules about how probability should be used. According to the
definition given above, it is legitimate to write:
P(L≤h≤U)=95%P(L≤h≤U)=95%
if (L, U) is derived according to a rule so that it does contain h for 95% of samples. This
resembles a subjective statement about our certainty that h lies in the range (L, U). But it
isn't: it's an objective statement about how often, in the long run, our random interval will
contain the fixed but unknown h, according to the randomness in our sampling. The
"subject" of the probability statement above looks like it is h (and that is the trick) but
actually it is the interval.

The probability statement above makes sense only before we draw a sample. What
happens after we draw the sample, when we find L=70 and U=74? There is a strong
temptation to plug into the previous expression to get:
P(70≤h≤74)=95%P(70≤h≤74)=95%
which is exactly the sentence held to be meaningless by frequentists earlier. h either lies
between the other two numbers or not; there is no probability involved.

Any inference we draw about h must of course happen after we draw the sample, but
frequentist rules prevent us from invoking probability at this point. So instead we refer back
to the randomness which gave us the interval (70, 74). It is not that this particular interval
contains h with 95% probability, but rather that an interval constructed in this way will
contain h 95% of the time

Everyone who uses confidence intervals, including every frequentist statistician on the
planet, would actually interpret the interval (70, 74) as representing a "likely range"
for h, implicitly invoking something like the illegal probability
statement P(70≤h≤74)=95%P(70≤h≤74)=95%. Without using some terms related to
probability, it is almost impossible to explain what useful relation the interval (70, 74) bears
to h.

Confidence intervals are useful mainly because we can misinterpret them as probability
statements about unknown parameters.

5. How would you explain to a group of senior executives why data is important?

• Examples