Recommender Systems Quiz

Practice Quiz: Recommender Systems Implementation

Question 1

Lecture described using ‘mean normalization’ to do feature scaling of the ratings. What equation below best describes this algorithm?

Option 1:

• y_norm(i,j) = (y(i,j) - μᵢ) / σᵢ where μᵢ = (1/Σⱼr(i,j)) * Σⱼ:r(i,j)=1 y(i,j) and σᵢ² = (1/Σⱼr(i,j)) * Σⱼ:r(i,j)=1 (y(i,j) - μⱼ)²

Option 2:

• y_norm(i,j) = y(i,j) - μᵢ where μᵢ = (1/Σⱼr(i,j)) * Σⱼ:r(i,j)=1 y(i,j) ✓

Option 3:

• y_norm(i,j) = (y(i,j) - μᵢ) / (maxᵢ - minᵢ) where μᵢ = (1/Σⱼr(i,j)) * Σⱼ:r(i,j)=1 y(i,j)

Mean normalization simply subtracts the average rating for each movie from the individual ratings. It doesn’t involve standard deviation (option 1) or min-max scaling (option 3).

Note

Reference: From section 6 - “Instead of using these original 0 to 5 star ratings over here, I’m going to take this and subtract from every rating the mean rating that it was given.”

---

Question 2

The implementation of collaborative filtering utilized a custom training loop in TensorFlow. Is it true that TensorFlow always requires a custom training loop?

• Yes. TensorFlow gains flexibility by providing the user primitive operations they can combine in many ways.

• No: TensorFlow provides simplified training operations for some applications. ✓

TensorFlow offers both approaches: simplified model.compile() and model.fit() for standard neural networks, and custom training loops for specialized algorithms like collaborative filtering.

Note

Reference: From section 7 - “If the model you have is a sequence of dense neural network layers or other types of layers supported by TensorFlow, and the old implementation recipe of model compound model fit works.”

---

Question 3

Once a model is trained, the ‘distance’ between features vectors gives an indication of how similar items are.

The squared distance between vectors x^(k) and x^(i) is: distance = ||x^(k) - x^(i)||² = Σₗ₌₁ⁿ (xₗ^(k) - xₗ^(i))²

Using the table below, find the closest item to “Pies, Pies, Pies”:

Movie	x₀	x₁	x₂
Pastries for Supper	2.0	2.0	1.0
Pies, Pies, Pies	2.0	3.0	4.0
Pies and You	5.0	3.0	4.0

Calculations:

• Distance to “Pastries for Supper”: (2.0-2.0)² + (2.0-3.0)² + (1.0-4.0)² = 0 + 1 + 9 = 10
• Distance to “Pies and You”: (5.0-2.0)² + (3.0-3.0)² + (4.0-4.0)² = 9 + 0 + 0 = 9

Answer: Pies and You ✓

“Pies and You” has a smaller squared distance (9) compared to “Pastries for Supper” (10), making it more similar to “Pies, Pies, Pies”.

Note

Reference: From section 8 - “The way we determine what are known as similar to the feature x^(i) is as follows: is the sum from l equals 1 through n with n features of x^(k)_l minus x^(i)_l square.”

---

Question 4

Which of these is an example of the cold start problem? (Check all that apply.)

• A recommendation system is unable to give accurate rating predictions for a new user that has rated few products. ✓

• A recommendation system is unable to give accurate rating predictions for a new product that no users have rated. ✓

• A recommendation system takes so long to train that users get bored and leave.

• A recommendation system is so computationally expensive that it causes your computer CPU to heat up, causing your computer to need to be cooled down and restarted.

The cold start problem specifically refers to the difficulty of making accurate predictions when there’s insufficient data - either for new users with few ratings or new items with few ratings.

Note

Reference: From section 8 - “One of this weaknesses is that is not very good at the cold start problem. For example, if there’s a new item in your catalog, say someone’s just published a new movie and hardly anyone has rated that movie yet” and “Similarly, for new users that have rated only a few items.”