Scribd
Upload
21 views

Netflix Data Science Interview Question

Data science

Uploaded by

Valeria Farias
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
21 views

Netflix Data Science Interview Question

Data science

Uploaded by

Valeria Farias
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

NETFLIX

DATA SCIENCE
INTERVIEW
QUESTIONS
WHAT IS DIFFERENCE BETWEEN BATCH AND ONLINE GRADIENT DESCENT
In batch gradient descent, the model looks at the entire dataset at
once to calculate the gradient (the direction to minimize error) and
update parameters. This means it calculates the average gradient
across all data points and then makes one update.
Pros: More stable and accurate, since it uses all data at each step.
Cons: Can be slow and memory-intensive, especially with large
datasets, since it needs to process all data at once.
Online Gradient Descent

In online gradient descent, the model updates parameters one data


point at a time. Each new data point provides a quick update
without waiting for all data to be processed. This approach is also
called stochastic gradient descent (SGD) because each update
uses a random single point, adding some randomness to the
updates.
Pros: Faster and requires less memory, as it only looks at one data
point at a time. Works well for very large datasets or data that’s
continuously updating (e.g., real-time applications).
Cons: Less stable because updates are noisier (one point at a time
can vary a lot), so it may "zigzag" toward the minimum rather than
taking a direct path.

@karunt
WHAT MAKE RELU AN EFFECTIVE ACTIVATION FUNCTION?

The Rectified Linear Unit, or ReLU works so well because -

Simplicity: ReLU is easy to compute. It simply takes any negative


value and turns it into zero, while keeping positive values as they are.
This makes it fast and efficient.
Formula: ReLU(x)=max(0,x)
ReLU(x)=max⁡(0,x)\text{ReLU}(x) = \max(0, x)

Avoids the Vanishing Gradient Problem: Many activation functions


(like sigmoid or tanh) squish values to be between -1 and 1, causing
gradients to shrink and slowing down learning. ReLU avoids this by not
limiting the positive side, allowing gradients to stay larger, which helps
in learning.

Sparse Activation: ReLU turns off (outputs zero) for any negative
input. This makes the network "sparse" by reducing unnecessary
signals, which improves efficiency and reduces the chances of
overfitting.

In short, ReLU is popular because it’s fast to compute, helps with


efficient learning, and prevents certain problems other functions have.

@karunt
EXPLAIN ANOVA TEST? (FOLLOW-UP: EXPLAIN MEANING OF P-VALUES)
The ANOVA test is used to compare the means of three or more groups
to determine if there is a significant difference among them. How ANOVA
Works -
1. Null Hypothesis (H₀): All group means are equal (i.e., any observed
differences are due to random variation).
2. Alternative Hypothesis (H₁): At least one group mean is significantly
different from the others.

A p-value represents the probability of observing the test results, or


something more extreme, under the assumption that the null hypothesis
is true.
Low p-value (< 0.05): Indicates that the observed data is unlikely
under the null hypothesis, so we have evidence to reject it in favor of
the alternative hypothesis. This suggests a statistically significant
effect.
High p-value (≥ 0.05): Suggests that the observed data is plausible
under the null hypothesis, so we fail to reject it. There isn’t strong
evidence for a significant difference.

@karunt
WHAT ARE THE KEY METRICS YOU WOULD CONSIDER WHEN EVALUATING
THE PERFORMANCE OF A RECOMMENDATION ALGORITHM?
Precision@K and Recall@K: These measure the relevance of the top K
recommended items. Precision@K is the proportion of relevant items in
the top K recommendations, while Recall@K measures the proportion of
all relevant items that appear in the top K.

Hit Rate: Measures how often the recommended list contains at least
one item that the user interacts with or rates highly, indicating that the
model is generating some relevant suggestions.

Normalized Discounted Cumulative Gain (NDCG): Measures the quality


of ranked lists by considering the position of relevant items in the
recommendations, with higher rewards for higher-ranking relevant
items. This is especially useful for ordered lists, like search results or top
recommendations.

Diversity: Evaluates how different the recommended items are from


each other. High diversity ensures users are not just shown similar items
repeatedly, which improves engagement.

Click-Through Rate (CTR): The ratio of users who click on


recommended items to those who view them. A higher CTR suggests
that the recommendations are capturing user interest.

@karunt
HOW WOULD YOU BUILD AND TEST A METRIC TO COMPARE TWO USERS’
RANKED LISTS OF MOVIE/TV SHOW PREFERENCES?
A few metrics to consider are -

Kendall’s Tau: Measures how similarly two lists are ranked by counting
the number of pairwise swaps needed to convert one list into the other.
It’s a good choice when you want to assess the order of preferences
rather than exact placement.

Spearman’s Rank Correlation: Measures correlation based on rank,


ignoring exact scores but comparing the relative order of items. It’s
helpful if you only have the ranks of items and want a measure that
handles ties.

Normalized Discounted Cumulative Gain (NDCG): This is a relevance-


based metric that measures how well a ranked list aligns with a ground
truth list, useful if some items in the list are considered more “relevant”
than others.

A few other things to consider are - making sure the list lengths are the
same, and conducting testing by swapping order to see how much the
above metrics change etc.

@karunt
WAS THIS HELPFUL?
Be sure to save it so you
can come back to it later!

@karunt

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy