This repository contains typical technical questions asked in data science interviews. Most of them are asked in interviews of Iran's big companies, but other famous questions found on the internet are also included.
After you find yourself ready to take an interview, reviewing these questions could be a great help. Since most of these questions are gathered from people's experience, it's recommended to answer them as if they were asked in the interview, e.g., answering with loud sound (not mumbling), giving the best possible answer, etc.
Note that most of the answers are given in the interview, so they might not be entirely accurate. You are encouraged to think about the questions and prepare your best possible solution.
If you've taken an interview in the past and would like to add any questions here, please append your question to the end of the list using the following format and beware that your kind support is enormously appreciated :).
#### <How old are you?>
- **Added by**: <John Abbasi> (you could also link you github account)
- **Interviewer Company**: <Tesla>
<Since I was born, the Earth has revoluted 29 times around the Sun.>
[⬆ Back to top](#questions)
Note that The sections wrapped in <> should be changed by the author, and other parts should remain the same. Please preview your README.md file to make sure your question doesn't break the style. Also, the two bullet points are optional. Here is what a question and answer would look like:
- Added by: John Abbasi
- Interviewer Company: Tesla
Since I was born, the Earth has revoluted 29 times around the Sun.
Alternatively, you could write your question(s) as issues (without following the format), and I will add them to the list in the above format.
If you have any problems/suggestions regarding the repository or want to add/change a question's answer, you can either submit a new issue about your concern or email me at mzarei137@gmail.com.
I would highly appreciate any other form of contribution.
Imagine we have gathered the height of all the people in Tehran, how does the distribution look like?
Since every natural event roughly follows the normal distribution, we expect the distribution to have precisely one peak value, no/low skewness, and roughly symmetric. The more the population, the closer it will look like the normal distribution.
It would still follow the normal distribution but with a lower variance (since the data is more similar) and probably a higher mean (since we remove a big chunk of data with low heights).
What if we separate the heights by gender? How does the distribution of height of all the men in Tehran look like?
It still follows the normal distribution but with a different mean and variance.
Still normal but with a skewness (heavy tail) on the right side.
Basic linear algebra questions. I recommend watching Gilbert Strang's 2005 linear algebra course (MIT).
Each value in the covariance matrix represents the covariance of two random variables (like features in a dataset) or two elements of a random vector. Its main diagonal values are variances of those random variables or random elements.
We have trained two different binary classification models on a dataset; The first one has 70% accuracy, and the other has 90%. Can we say for sure that the second one is better? Why not? Which metric is better to use here?
Accuracy metric performs poorly on an imbalanced dataset, and since we don't know anything about the class distribution, we cannot say which one is the best. Other valuable metrics often used alongside accuracy are recall, precision, and F1-score. Note that with highly imbalanced datasets, ROC AUC can be misleading.
We currently have a dataset which suffers from heavy label imbalance. Can you name some methods to fix it?
- Giving different weight to labels based on its frequency.
- Generating data belonging to the infrequent label.
- Down-sampling the abundant label.
- And more...
The mean wil roughly stay the same but the overall std might change depending on the std of the added data. See this link for more info about the new overall std.
Imagine we have a binary classifier which is going to classify a person as either an assassin or a normal person ("True" class means assassin and "False" class means normal). Which classification threshold is acceptable?
Depending on the task, we should consider different threshold for different metrics. In this case, we definitely do not want any assassin to be classified as a normal person but mistaking a few individuals as assassins is alright. So our "False Negative" values should be the lowest possible (zero ideally since we don't want anyone to be killed!).
In the assassin classification case, increasing the threshold will result in a bigger precision since we remove instances from the "False Positive" category.
What is the confusion matrix? What are the ROC curve and ROC AUC? What does each point on the ROC curve represent?
You can answers online since these are basic questions.
What is the model variance score? What is regularization, and what is it used for? Which one of the L1 and L2 regularizations is used in neural networks?
In general, regularization is a set of techniques used to overcome overfitting in different models. Both of them are used in NNs, but L2 regularization is more common.
What is PCA? What does each column in PCA results represent? Could we say "Column 1 of PCA represents the Age"?
PCA. Since PCA tries to create new features from the composition of some features, the statement won't be correct in general.
What happens when we make the cosine distance our distance metric (and not the euclidian distance) in clustering?
- Plot the inertia score vs. the number of clusters, and choose the "elbow" number.
- (More precise but more computationally expensive) For each number of clustering, calculate the "silhouette score" (the mean "silhouette coefficient" over all instances), and choose the number that maximizes the score.
Let's say we have chosen a specific number of clusters and we want to see how good is the model. How could we evaluate the clustering based on the selected number of clusters?
In general, we want the distance between instances in a cluster to be the lowest and the distance between points clusters to be the highest.
- Improving the style.
- Adding individual answers to a question.