Week 8 · Assessment Quiz

Collaborative Filtering

20 multiple-choice questions on core concepts, plus 5 short-answer questions on FastAI. Attempt every question before submitting.

📋 25 questions total ⭐ 25 marks 🕐 No time limit 🔒 Answers not revealed

Collaborative filtering makes recommendations based on:

AThe genre and cast metadata attached to each item BThe rating patterns shared across many users and items CThe demographic profile of the individual user (age, location) DThe most recently released items in the catalogue

In a user–item rating matrix, a missing cell (marked ? or NaN) means:

AThe user gave the item a rating of zero BThe item was removed from the catalogue CThe user has not yet interacted with that item — the rating is unknown DThe rating was corrupted and should be dropped from the dataset

A rating matrix with 500 users, 1,000 movies, and only 10,000 known ratings has a sparsity of:

A2% (matrix is 2% empty) B98% (matrix is 98% empty) C50% (matrix is half empty) D10% (matrix is 10% empty)

In matrix factorisation, the rating matrix R is approximated as:

AR ≈ U + V (element-wise addition of two matrices) BR ≈ U × Vᵀ (product of a user matrix and a transposed item matrix) CR ≈ U ÷ V (element-wise division of two matrices) DR ≈ log(U) × V (logarithm of the user matrix times the item matrix)

The parameter k in matrix factorisation refers to:

AThe number of training epochs BThe total number of known ratings in the dataset CThe number of latent dimensions in each user and item embedding vector DThe learning rate used during gradient descent

A predicted rating in the basic dot-product CF model is computed as:

AThe Euclidean distance between the user and item embedding vectors BThe sum of all values in the user embedding vector CThe dot product of the user embedding vector and the item embedding vector DThe average of the user's past ratings

Latent factors in collaborative filtering are best described as:

AHand-crafted features like genre, director, and release year BHidden preference dimensions discovered automatically from rating patterns CThe number of users who rated each item DThe weights in the final output layer of a neural network

A user bias term b_u captures:

AHow much the model overfits to a particular user's ratings BThe tendency of a user to rate all items systematically higher or lower than average CThe number of items the user has rated so far DThe difference between the user's first and last rating

Given user embedding u = [0.8, 0.5] and item embedding v = [0.6, −0.4], what is the dot product u · v?

A0.68 B0.28 C0.48 D0.96

Q10

The loss function used to train a collaborative filtering model on known ratings is typically:

ACross-entropy loss BMean Absolute Error (MAE) CMean Squared Error (MSE) over known cells only DHinge loss

Q11

A model predicts [4.2, 3.1, 2.8] for three known ratings of [4, 3, 3]. What is the RMSE?

Aapproximately 0.17 Bapproximately 0.42 Capproximately 0.05 Dapproximately 0.73

Q12

Weight decay (L2 regularisation) is applied during CF training primarily to:

ASpeed up convergence by scaling gradients BPrevent embedding values from growing too large and overfitting CEnsure the embeddings sum to exactly 1.0 DReduce the number of latent factors automatically

Q13

The cold start problem in collaborative filtering occurs when:

AThe training dataset is too large to fit in memory BA new user or item has no historical ratings, so the model cannot personalise recommendations CThe model is run on a machine without a GPU DThe learning rate is set too low and training never converges

Q14

PCA is used to visualise learned embeddings because:

APCA improves the accuracy of predictions before plotting BIt reduces high-dimensional embedding vectors to 2D so they can be plotted on a scatter chart CPCA removes noisy ratings from the training data DIt ensures all embedding values lie between 0 and 1

Q15

Cosine similarity between two embedding vectors measures:

AThe straight-line distance between the two points in embedding space BThe angle between the two vectors — how similar their directions are, regardless of magnitude CThe sum of squared differences between corresponding elements DThe number of elements the two vectors have in common

Q16

Neural collaborative filtering (with use_nn=True) differs from dot-product CF because it:

ADoes not use any embedding vectors at all BConcatenates the user and item embeddings and passes them through an MLP instead of computing a dot product CReplaces gradient descent with a closed-form solution DUses convolutional layers to extract patterns from the rating matrix

Q17

Why is y_range=(0, 5.5) preferred over (0, 5.0) when ratings go from 1 to 5?

ATo allow the model to predict ratings above 5 for exceptional items BTo avoid sigmoid saturation at the boundaries, which would cause gradients to vanish and learning to stall CBecause 5.5 is the standard maximum in the MovieLens dataset DTo add a 0.5-star tolerance for measurement error in user ratings

Q18

The filter bubble effect in recommender systems refers to:

AA technical bug that causes duplicate recommendations to appear BThe phenomenon where CF reinforces existing preferences, narrowing the diversity of content users are exposed to CA memory error that occurs when the rating matrix is too large DThe tendency of new items to be recommended more than established ones

Q19

A model has 800 users, 400 movies, and uses k=50 latent factors. How many embedding parameters does it have (excluding biases)?

A40,000 B16,000,000 C60,000 D600

Q20

Which of the following is a valid way to handle the cold start problem for a new user?

AAssign the new user the same embedding as the most active existing user BAsk the new user to rate a few items during onboarding to bootstrap their embedding CRemove all new users from the dataset until they have rated at least 100 items DIncrease the learning rate specifically for new user embeddings

Answer each question in 1–3 sentences using correct FastAI terminology. These questions are marked by your lecturer. Aim for precision — code snippets where relevant are encouraged.

Q21

What FastAI function is used to create a DataLoader for collaborative filtering, and what are the three column names you must specify? written

Your answer

Hint: think about what describes a single rating row.

0 / 600

Q22

After training a collab_learner model, calling learn.model.forward() directly causes a TypeError. What is the correct FastAI way to generate predictions for new user–item pairs? written

Your answer

Hint: involve dls.test_dl() and learn.get_preds().

0 / 600

Q23

dls.n_users and dls.n_items raise an AttributeError in current versions of FastAI. How do you correctly retrieve the number of unique users and items from a CollabDataLoaders object? written

Your answer

Hint: use dls.classes.

0 / 600

Q24

What does the use_nn=True argument do when passed to collab_learner()? How does the resulting architecture differ from the default dot-product model? written

Your answer

0 / 600

Q25

Explain what learn.model.i_weight.weight and learn.model.i_bias.weight contain after training, and describe one way you could use each to interpret what the model has learned. written

Your answer

0 / 600

Your full name

Check that you have answered all 25 questions, then click Submit. Your MCQ score (out of 20) will be shown. Short-answer questions are marked separately by your lecturer.

Collaborative Filtering

Multiple Choice (1 mark each)

Short Answer — FastAI (1 mark each)