Question 1

Imagine that we apply X = PCA(n_components=5).fit_transform(data) and data has shape (5000, 53). What is the shape of X?

Correct answers:

• (5000, 5) . Yes, it should be (n_samples, n_components).

Incorrect answers:

• (5, 53) . No, we cannot transform 5000 samples into 5.

• (5, 5000) . No, we cannot transform 5000 samples into 5.

• (53, 5) . No, we cannot transform 5000 samples into 53.

Question 2

To which data NMF is NOT applicable?

Correct answers:

• Standartized matrix. "Standartized" means that every feature column has zero mean and unit variance. This implies that we have negative values and cannot apply NMF.

Incorrect answers:

• Bag-of-words matrix. Since BoW matrix is non-negative matrix, we *can* apply NMF to it.

• One-Hot encoded feature. Since this matrix contains only 0's and 1's -- it is non-negative and we can apply NMF.

Question 3

Suppose we have 2 categorical features: f1 with A possible values and f2 with B possible values. How many values will their interaction have?

Correct answers:

• Less or equal to A * B. True. Sometimes some value (e.g. *a*) from A cannot be used with some value (eg *b*) from B. In this case, we have no change to see *ab* combination. If all value from A can be used with all values from B -- we will get A*B new possible values.

Incorrect answers:

• Exactly A + B. No, it's too small.

• Exactly A * B. Not exactly, sometimes some value (e.g. *a*) from A cannot be used with some value (eg *b*) from B. In this case, we have no change to see *ab* combination.

• max(A, B). No, it's way too small.

Question 4

Imagine we have 2 categorical features represented as integers: f1 with all values in range [0, 1000] and f2 with values in range [0, 100]. What is the correct way to build their interaction?

Correct answers:

• f1.astype(str) + "_" + f2.astype(str). Yes, this is the right answer

Incorrect answers:

• f1 + f2 . There are some problems. For example, if f1+f2=100: it is an interaction of 0 and 100, 100 and 0, or 90 and 10?

• f1.astype(str) + f2.astype(str). There is some problems. For example, "123": it is an d interaction of "1" and "23" or "12" and "3"?

• (f1 + f2).astype(str). It is essentially the same way as just f1+f2.

Question 5

What is a correct way to get t-SNE projection of train and test data?

Correct answers:

• Apply t-SNE to concatenation of train and test and split projection back. This this the rigth way since train and test will projected in the same way.

Incorrect answers:

• Apply t-SNE to the train and after that to the test. This this case we will have 2 different projections.

• Apply t-SNE to the test first and after to train. This this case we will have 2 different projections.

• Doesn't matter, all variants will produce the same result. No, since tSNE results in randomized projection

Question 6

Is it possible to do t-SNE projection into 20-dimensional space?

Correct answers:

• Yes, why not. You can do tSNE projection into arbitrary space.

Incorrect answers:

• No, only 2-dim or 3-dim projections are possible. Wrong! Despite tSNE is quite often used for visualization purposes, you don't limit to use only 2- or 3-dimentional projections.