WorksheetsBA 10
Total questions: 72
Worksheet time: 36mins
Name
Class
Date
1.
According to the lecture, what shift do recommender systems enable to address information overload?
a)
From passive discovery to active search
b)
From active search to passive discovery
c)
From batch processing to stream processing
d)
From supervised learning to unsupervised learning
e)
From personalization to standardization
2.
In the long tail effect, what is a key goal of a recommender system?
a)
Maximize sales of only the most popular items
b)
Help users discover niche items and increase catalog utilization
c)
Eliminate niche items to simplify choice
d)
Reduce the number of items shown to users permanently
e)
Optimize inventory replenishment for warehouses
3.
Which data source is an example of implicit feedback?
a)
A 1 to 5 star rating
b)
A written product review
c)
Watch time on a video
d)
A survey response about preferences
e)
A like or dislike button click
4.
Collaborative filtering primarily focuses on which information to make recommendations?
a)
Item attributes and metadata
b)
User interactions and behavior patterns
c)
Random exploration only
d)
Manual expert rules
e)
Price elasticity estimates
5.
In the MovieStream case study, what is the stated objective?
a)
Cluster movies into genres automatically
b)
Predict which movie Alice will watch next
c)
Detect fraudulent user accounts
d)
Compress movie trailers for streaming
e)
Estimate the studio budget of each movie
6.
In content-based recommendation, why is an item recommended to a user?
a)
It is the most popular item overall
b)
It is very similar in features to an item the user liked
c)
It has the highest price among candidates
d)
It was released most recently
e)
It was clicked by the largest number of users
7.
Which technique is used to convert plot summaries and descriptions into a keyword-weighted vector for item profiles?
a)
One-hot encoding
b)
TF-IDF
c)
K-nearest neighbors
d)
Decision trees
e)
Principal component analysis
8.
How is a user profile constructed in the content-based approach described in the slides?
a)
A random vector assigned at signup
b)
A weighted average of the vectors of items the user consumed
c)
A list of the top 10 most popular items
d)
A count of how many users watched each item
e)
A manual set of rules written by domain experts
9.
A cosine similarity of 1.0 between a user profile vector and an item vector indicates:
a)
The vectors are orthogonal and unrelated
b)
A perfect match because the vectors align
c)
The item is guaranteed to be unpopular
d)
The user has no interaction history
e)
The system should assign zero confidence
10.
In the content-based workflow, what does candidate generation do?
a)
Select only items the user has already seen
b)
Collect all items not yet seen by the user as candidates
c)
Replace user vectors with item vectors
d)
Remove all niche items from consideration
e)
Sort items by release year only
11.
Which benefit of content-based filtering helps with brand-new items?
a)
It does not require any item metadata
b)
It can recommend a new item immediately if metadata exists
c)
It requires large numbers of ratings for every new item
d)
It guarantees high diversity by design
e)
It eliminates the need for feature engineering
12.
Which drawback of content-based filtering can reduce serendipity?
a)
High diversity by default
b)
Overspecialization leading to a filter bubble
c)
Perfect handling of new users with no history
d)
Automatic correction of noisy clicks
e)
No need to tag items
13.
What core assumption underlies collaborative filtering in the lecture?
a)
Items with similar titles are always similar
b)
Users who agreed in the past will agree in the future
c)
Only explicit ratings can be used for recommendations
d)
Metadata is more important than behavior
e)
Users never change preferences over time
14.
Why does the lecture highlight item-based collaborative filtering as often preferred over user-based?
a)
Users are more stable than items
b)
Items are more stable and fewer calculations are needed in real time
c)
It eliminates sparsity in the interaction matrix
d)
It works without any interaction data
e)
It requires no model training
15.
What is the main structural challenge of the user-item interaction matrix described in the slides?
a)
It is always dense and full of values
b)
It is highly sparse with most entries missing
c)
It contains only text fields
d)
It cannot represent users and items together
e)
It has no rows and no columns
16.
When dealing with implicit data, what modeling idea is introduced to handle ambiguity and noise?
a)
Treat every missing value as a dislike
b)
Separate preference from confidence
c)
Remove all users with few interactions
d)
Convert all interactions to 1 to 5 star ratings
e)
Use only content-based filtering
17.
In the implicit feedback model, how is confidence defined from an interaction count r_ui?
a)
c_ui = alpha - r_ui
b)
c_ui = 1 + alpha * r_ui
c)
c_ui = r_ui / alpha
d)
c_ui = log(alpha + r_ui) only
e)
c_ui is always 1 for all interactions
18.
Matrix factorization in the lecture approximates the interaction matrix R by:
a)
A single decision tree
b)
A product of two smaller dense matrices U and V transpose
c)
A list of nearest neighbors only
d)
A set of hand-crafted rules
e)
A correlation matrix of item metadata
19.
Latent factors are best described as:
a)
Visible genre labels assigned by editors
b)
Hidden features learned by the model
c)
Exact user demographics collected at signup
d)
The raw click counts r_ui
e)
A list of all items in the catalog
20.
How is a user's predicted interest score for an item computed in matrix factorization?
a)
Sum of item prices
b)
Dot product of the user vector and the item vector
c)
Difference between the user ID and item ID
d)
Average of all users' ratings
e)
Random sampling from the long tail
21.
What is the core training loop of Alternating Least Squares (ALS) described in the lecture?
a)
Update both user and item vectors simultaneously in one step
b)
Fix user vectors and solve for item vectors, then fix item vectors and solve for user vectors
c)
Only update user vectors and never item vectors
d)
Only update item vectors and never user vectors
e)
Train a neural network with backpropagation only
22.
In the simplified example, why is the action movie recommended to Alice instead of the romance movie?
a)
Its dot product score with Alice's vector is higher
b)
It has the newest release date
c)
It has the highest production budget
d)
It has the lowest similarity to Alice's profile
e)
It is the least watched item overall
23.
Which is listed as an advantage of collaborative filtering?
a)
It always requires item metadata
b)
It increases serendipity and helps discovery
c)
It cannot be used outside movies
d)
It is immune to cold-start issues
e)
It guarantees the best RMSE
24.
Popularity bias in collaborative filtering refers to:
a)
A preference for niche items only
b)
A tendency to recommend items that are already popular
c)
A guarantee that every item will be recommended equally
d)
A requirement for explicit star ratings
e)
A method for removing noise from clicks
25.
Which statement correctly contrasts content-based vs collaborative filtering for new items?
a)
Content-based is hard for new items while collaborative filtering is easy
b)
Content-based can recommend new items if metadata exists, while collaborative filtering struggles due to cold-start
c)
Both methods require millions of explicit ratings for new items
d)
Collaborative filtering can recommend new items without any interactions, while content-based cannot
e)
Neither method can ever recommend new items
26.
Why does the lecture distinguish offline and online evaluation for recommenders?
a)
Offline uses historical train-test splits, while online uses live A/B testing for business impact
b)
Offline is only for images, while online is only for text
c)
Offline requires no data, while online requires no data
d)
Offline always replaces online evaluation
e)
Online evaluation ignores user behavior
27.
For implicit feedback, which metric category is emphasized as crucial?
a)
Prediction accuracy metrics like RMSE only
b)
Ranking accuracy metrics like Precision, Recall, and NDCG
c)
Compression ratio metrics
d)
Network latency metrics
e)
Feature importance metrics
28.
Why can a model with good RMSE still produce a poor user experience, according to the lecture?
a)
Users only care about the exact decimal rating values
b)
Users care about ranking order more than the exact rating values
c)
RMSE cannot be computed from explicit ratings
d)
RMSE is only an online metric
e)
RMSE automatically guarantees high novelty
29.
In top-N evaluation, what is a 'hit'?
a)
A recommended item that matches an item in the user's test set consumption
b)
Any item shown on the homepage
c)
An item with the highest price
d)
An item with zero interactions
e)
A user who never clicks
30.
According to the slides, when the cut-off K increases, what is the expected precision-recall trade-off?
a)
Both precision and recall decrease
b)
Precision increases while recall decreases
c)
Recall increases while precision decreases
d)
Both precision and recall remain unchanged
e)
Precision becomes equal to recall
31.
Recall@K is intended to measure:
a)
The percent of recommendations that are relevant
b)
The percent of relevant items found by the system
c)
The average rating error
d)
The number of users in the system
e)
The time it takes to generate recommendations
32.
Mean Average Precision (MAP) is designed to:
a)
Ignore the position of relevant items
b)
Reward systems that rank relevant items near the top by averaging precision across ranks
c)
Measure only business revenue
d)
Replace A/B testing completely
e)
Compute cosine similarity between items
33.
NDCG discounts relevance primarily based on:
a)
Whether the user is new or old
b)
The rank position in the recommended list
c)
The number of genres an item has
d)
The item release year
e)
The total catalog size
34.
In online A/B testing, what distinguishes Group A from Group B in the setup described?
a)
Group A is the treatment and Group B is the control
b)
Group A uses the existing algorithm (control) and Group B uses the new algorithm (treatment)
c)
Both groups always see identical recommendations
d)
Group A is only for offline evaluation
e)
Group B is only for explicit ratings
35.
Which KPI is most directly associated with engagement in recommendations?
a)
CTR (Click-Through Rate)
b)
MAE (Mean Absolute Error)
c)
RMSE (Root Mean Squared Error)
d)
NDCG (Normalized Discounted Cumulative Gain)
e)
TF-IDF (Term Frequency-Inverse Document Frequency)
36.
Which approach beyond matrix factorization explicitly incorporates signals like time or location?
a)
Content-based filtering
b)
Context-aware recommenders
c)
User-user nearest neighbors only
d)
One-hot encoding only
e)
RMSE optimization only
37.
Which Python library listed in the slides is optimized for implicit feedback and ALS?
a)
Surprise
b)
Implicit
c)
TensorFlow Recommenders
d)
LightFM
e)
NumPy
38.
Which statement matches a key takeaway about implicit feedback from the lecture?
a)
Implicit feedback can be treated exactly like explicit star ratings without changes
b)
Implicit feedback requires special handling using confidence weighting
c)
Implicit feedback is always unbiased and noise-free
d)
Implicit feedback removes the need for evaluation metrics
e)
Implicit feedback guarantees no cold-start issues
39.
A new movie has rich genre and plot metadata but no user interactions yet. Which approach from the lecture is most likely to recommend it immediately, and why?
a)
Collaborative filtering, because it needs crowd behavior only
b)
Content-based filtering, because it uses item attributes
c)
User-user nearest neighbors, because it needs ratings history
d)
A/B testing, because it generates recommendations directly
e)
RMSE optimization, because it creates metadata
40.
A brand-new user has not watched or rated anything. Based on the lecture, what is the most accurate statement about cold-start for this user?
a)
Content-based works perfectly because metadata alone is enough
b)
Collaborative filtering works perfectly because the community is large
c)
Both content-based and collaborative filtering struggle because there is no user history
d)
Only matrix factorization works because it ignores history
e)
Cold-start is only a problem for new items, not new users
41.
A platform mostly recommends blockbusters and rarely surfaces niche titles. Which concept best explains what is being missed, and what is the intended RecSys goal?
a)
Long tail; increase catalog utilization by helping users discover niche items
b)
Cold-start; remove all new items from the catalog
c)
RMSE; minimize rating error for every item
d)
CTR; maximize clicks by showing only popular items
e)
TF-IDF; reduce text dimensionality for speed
42.
Your interaction logs contain accidental clicks and varying watch times. In the lecture's implicit feedback framing, which interpretation is most appropriate?
a)
Preference is binary, and confidence increases with interaction intensity
b)
Preference is the exact watch time in minutes, and confidence is always zero
c)
Preference is always negative for missing entries, and confidence is fixed at 1
d)
Preference comes only from explicit star ratings, and confidence comes from comments
e)
Preference is random, and confidence is determined by item price
43.
If you treat every missing interaction in an implicit dataset as a negative preference, which lecture point are you contradicting?
a)
Missing data is always a dislike signal
b)
Missing data is not the same as dislike in implicit feedback
c)
Implicit feedback contains no noise
d)
Implicit feedback is only for explicit ratings
e)
Implicit feedback makes sparsity disappear
44.
You want recommendations that can be explained as 'Recommended because you watched X'. Which advantage and approach from the lecture align best with that requirement?
a)
Transparency; content-based filtering
b)
Serendipity; content-based filtering
c)
No item cold-start; collaborative filtering
d)
Popularity bias; collaborative filtering
e)
A/B testing; matrix factorization
45.
A company wants neighborhood-based collaborative filtering but needs fewer real-time computations. Which variant does the lecture motivate, and why?
a)
User-based, because users are more stable than items
b)
Item-based, because items are more stable and require fewer real-time calculations
c)
User-based, because it eliminates sparsity
d)
Item-based, because it does not need interactions
e)
Hybrid, because it removes the need for ranking
46.
In matrix factorization, users and items live in the same latent space. Operationally, when is an item considered a good recommendation for a user?
a)
When the user and item vectors have a high dot product (aligned directions)
b)
When the user and item vectors are orthogonal
c)
When the item has the most metadata tags
d)
When the item has the lowest price
e)
When the item has zero interactions
47.
Why does the lecture argue standard gradient descent is difficult for matrix factorization with implicit data?
a)
There are too many zeros, making optimization challenging in the implicit setting
b)
Gradient descent cannot run on modern hardware
c)
Implicit data always forms a dense matrix
d)
Implicit data provides perfect negative labels
e)
ALS is slower than gradient descent
48.
A model improves RMSE on explicit ratings, but users still complain the recommendations feel worse. Which lecture point best explains how this can happen?
a)
Users care more about ranking order than exact rating decimals, so good RMSE may not imply good experience
b)
RMSE is only defined for implicit feedback
c)
A lower RMSE always means a better ranked list
d)
User complaints are irrelevant to evaluation
e)
RMSE directly measures CTR
49.
For the same system, you increase the cut-off K from 10 to 50 in top-N evaluation. What trade-off should you expect, according to the lecture?
a)
Recall tends to increase while precision tends to decrease
b)
Precision and recall both increase
c)
Precision tends to increase while recall tends to decrease
d)
Both precision and recall become identical
e)
Recall and precision are unaffected by K
50.
Two systems have the same Precision@10, but System A places relevant items earlier in the list. Which offline metric from the lecture is introduced to reward earlier hits in a ranked list?
a)
RMSE
b)
Recall@10
c)
CTR
d)
Mean Average Precision (MAP)
e)
MAE
51.
When relevance is graded (not just hit or miss), which metric is presented as the gold standard for ranking evaluation?
a)
RMSE
b)
MAE
c)
NDCG
d)
CTR
e)
TF-IDF
52.
Which statement best captures why offline evaluation is inherently challenging for recommender systems?
a)
We do not know whether users like items they have not seen, so historical data provides incomplete ground truth
b)
Offline evaluation always has perfect labels for all items
c)
Offline evaluation measures business impact directly
d)
Offline evaluation requires no train-test split
e)
Offline evaluation avoids sparsity automatically
53.
In the A/B test setup described, what is the correct meaning of the treatment group?
a)
Users who never click
b)
Users shown the existing algorithm
c)
Users shown the new algorithm being tested
d)
Users who only rate items explicitly
e)
Users who are excluded from analysis
54.
A team wants to prevent boredom from showing many very similar items. Which KPI category from the lecture most directly targets this concern?
a)
CTR
b)
CVR
c)
Diversity
d)
RMSE
e)
MAE
55.
A recommender uses a sequence of recent actions within a session to predict the next item. Which 'beyond matrix factorization' approach listed fits best?
a)
Neural Collaborative Filtering only
b)
RNNs or Transformers for session-based recommendation
c)
Cosine similarity on TF-IDF vectors only
d)
User-user memory-based CF only
e)
Offline train-test split only
56.
A retailer has weak product descriptions but abundant purchase co-occurrence data. Which advantage of collaborative filtering is most relevant?
a)
It works without understanding item features (content agnostic)
b)
It removes the need for any user interactions
c)
It guarantees no cold-start issues
d)
It always produces transparent explanations
e)
It requires only item metadata
57.
If a recommender over-emphasizes already popular items, which downside and business objective from the lecture are most likely harmed?
a)
Popularity bias; long-tail discovery and catalog utilization
b)
Overspecialization; parallel training with ALS
c)
Noise; the ability to compute TF-IDF
d)
Cold-start; the ability to compute RMSE
e)
Diversity; the ability to compute dot products
58.
A team wants to reduce overspecialization from content-based methods and cold-start from pure collaborative filtering. According to the lecture taxonomy, what system class should they consider?
a)
Only content-based filtering
b)
Only user-based collaborative filtering
c)
Only item-based collaborative filtering
d)
Hybrid systems combining content-based and collaborative filtering
e)
Only A/B testing
59.
In the implicit feedback confidence model c_ui = 1 + alpha * r_ui, what is the most likely consequence of choosing an excessively large alpha?
a)
All confidence weights become nearly equal regardless of behavior
b)
High-count interactions dominate learning and can drown out rarer signals
c)
The interaction matrix becomes dense and no longer sparse
d)
Implicit feedback turns into explicit 1 to 5 star ratings automatically
e)
CTR is guaranteed to increase in online tests
60.
Content-based filtering can recommend a new item if metadata exists. Which failure mode becomes most severe when metadata is sparse or inconsistently tagged?
a)
Cold-start for new items becomes worse than collaborative filtering
b)
Feature engineering and tagging quality become a bottleneck for recommendation quality
c)
Matrix factorization cannot be trained anymore
d)
A/B testing becomes impossible
e)
The interaction matrix becomes less sparse
61.
A marketplace adds many new items daily and also wants high serendipity. Based on the lecture, which design best addresses both aims?
a)
Pure collaborative filtering only
b)
Pure content-based filtering only
c)
A hybrid that uses content for new items and collaborative signals for discovery
d)
Only RMSE optimization on explicit ratings
e)
Only CTR optimization without personalization
62.
Offline RMSE improves after an update, but an online A/B test shows lower watch time. Which explanation is most consistent with the lecture?
a)
RMSE captures business impact directly, so the A/B test must be wrong
b)
Offline metrics are proxies and may not reflect ranking quality or user experience
c)
If RMSE improves, CTR must always improve too
d)
A/B tests should never compare control and treatment
e)
Watch time cannot be measured for recommender systems
63.
A system optimizes Precision@K with a very small K and achieves high usefulness for the first few items, but users complain it misses many things they would like. What is the best diagnosis and the metric that would expose it?
a)
High recall; use RMSE to confirm
b)
Low coverage (low recall); use Recall@K to measure it
c)
High diversity; use CVR to measure it
d)
Low novelty; use TF-IDF to measure it
e)
High serendipity; use MAE to measure it
64.
Why is modeling preference as binary but adding confidence weighting a reasonable approach for implicit feedback, according to the lecture framing?
a)
Implicit logs provide perfect negative labels, so confidence is unnecessary
b)
Interaction indicates some positive signal, while intensity provides confidence about that signal
c)
Binary preference eliminates the need for user-item matrices
d)
Confidence weighting makes all items equally likely to be recommended
e)
Confidence weighting converts implicit data into explicit ratings
65.
A user watches one action movie repeatedly. With a weighted-average user profile and cosine similarity, what outcome is most likely under pure content-based filtering?
a)
The system increases diversity by recommending unrelated genres
b)
The system becomes more specialized toward that movie's features, increasing filter-bubble risk
c)
The system cannot update the user profile until retraining
d)
The system switches to collaborative filtering automatically
e)
The system treats repeated watches as negative feedback
66.
Item-based collaborative filtering is motivated partly by item stability. In which scenario does that advantage weaken the most?
a)
The user base grows while the catalog stays fixed
b)
Items and their relationships change rapidly (for example seasonal or frequently refreshed catalogs)
c)
Users provide more explicit star ratings
d)
The interaction matrix becomes more sparse
e)
The platform runs ALS on a single machine
67.
Two systems each recommend 10 items and both contain the same number of relevant items, but System A ranks the relevant items at positions 1 to 3 while System B ranks them at positions 8 to 10. Which metric from the lecture would most strongly prefer System A?
a)
Precision@10
b)
Recall@10
c)
RMSE
d)
Mean Average Precision (MAP)
e)
CTR
68.
You train matrix factorization using implicit watch-time logs but evaluate success using RMSE as if watch time were a 1 to 5 rating. What is the most critical mismatch highlighted by the lecture?
a)
RMSE is only defined for online evaluation
b)
Implicit feedback evaluation should focus on ranking metrics, and missing data is not a clear negative label
c)
Cosine similarity cannot be used with matrix factorization
d)
A/B tests do not apply to recommender systems
e)
Latent factors cannot exist in implicit data
69.
You want to increase long-tail discovery while monitoring business impact in an A/B test. Which set of online KPIs from the lecture is most informative to track together?
a)
CTR and CVR only, ignoring diversity and novelty
b)
RMSE and MAE only, ignoring user behavior
c)
CTR along with diversity and novelty to balance engagement and discovery
d)
TF-IDF and cosine similarity only
e)
User count and item count only
70.
Why do both a new user and a new item create severe cold-start for collaborative filtering in the lecture?
a)
Both lack interaction history needed to form neighbors or learn latent vectors
b)
Both have too much metadata, which confuses collaborative models
c)
Both automatically increase sparsity above 99 percent
d)
Both make cosine similarity undefined for all users
e)
Both eliminate the need for evaluation metrics
71.
Which property makes ALS particularly suitable for big data environments in the lecture?
a)
It guarantees perfect diversity
b)
It is highly parallelizable
c)
It removes the need for train-test splits
d)
It eliminates cold-start for new users
e)
It works without any interaction matrix
72.
A niche domain has strong item metadata but very few users, and leadership still wants novelty. Based on the lecture trade-offs, what is the best starting approach and the main limitation to mitigate?
a)
Start with collaborative filtering; mitigate the need for item features
b)
Start with content-based filtering; mitigate overspecialization and low serendipity
c)
Start with A/B testing; mitigate sparsity in the interaction matrix
d)
Start with RMSE optimization; mitigate the long tail effect
e)
Start with user-user CF; mitigate transparency
100 %
