WorksheetsWeb Mining 8
Total questions: 105
Worksheet time: 53mins
Name
Class
Date
1.
Why do users need recommender systems on the web?
a)
Users are overloaded with information
b)
Users always know exactly what they want
c)
The web has too few products to choose from
d)
Recommenders remove the need for any data
e)
Recommenders replace all search engines
2.
In information retrieval, how do users express their wishes?
a)
Through a query
b)
Through a random item click only
c)
By rating every item first
d)
By selecting a neighbor set
e)
By training a neural network manually
3.
Which is listed as an application area for recommender systems?
a)
Ecommerce
b)
Operating system kernels
c)
Database indexing only
d)
Compiler optimization
e)
Network routing
4.
According to the examples, which company uses recommendations for news suggestions?
a)
Google News
b)
Amazon
c)
Netflix
d)
Wikipedia
e)
Stack Overflow
5.
What is one reason to consider time in recommender systems?
a)
User preferences and behavior can change over time
b)
Time stamps are never available in datasets
c)
Time always reduces prediction accuracy
d)
Time removes the need for evaluation
e)
Time makes ratings identical across users
6.
Which method suggests items based on a user's transaction history?
a)
Content-based
b)
Collaborative filtering
c)
Session-based
d)
Random recommendation
e)
Manual curation only
7.
In the Netflix challenge context, what is the goal of improving the recommender algorithm?
a)
Reduce prediction error and improve recommendation accuracy
b)
Increase the number of items in the catalog
c)
Remove the need for user ratings
d)
Guarantee the same recommendations for everyone
e)
Replace evaluation metrics with intuition
8.
A chart showing more recommender-system papers over the years mainly indicates what?
a)
Growing research interest in recommender systems
b)
A decline in personalization needs
c)
That evaluation metrics are no longer used
d)
That collaborative filtering is obsolete
e)
That user profiles are unnecessary
9.
In a content-based approach, recommendations are based primarily on what?
a)
Item features and a user profile
b)
Only the global average rating
c)
Only the most recent click by any user
d)
Randomly chosen items
e)
The number of servers in a cluster
10.
Collaborative filtering mainly uses what type of data to recommend items?
a)
Community data from multiple users
b)
Only item descriptions
c)
Only product prices
d)
Only website layout information
e)
Only user passwords
11.
What is the cold-start challenge in recommender systems?
a)
Not enough information about new users or new items
b)
Too many ratings for every user-item pair
c)
Having only one item in the catalog
d)
Having perfect knowledge of user preferences
e)
Training always finishes instantly
12.
In evaluation notation, what does U represent?
a)
Set of users
b)
Set of items
c)
Set of timestamps
d)
Set of similarity functions
e)
Set of loss functions
13.
Which rating scale is mentioned as an example of explicit feedback?
a)
A 5-stage scale from 1 to 5
b)
A 0 to 1000 scale only
c)
A letter-grade scale only
d)
A three-color traffic-light scale
e)
No rating scale is mentioned
14.
Why is a dataset split into train and test sets in recommender evaluation?
a)
To train on one part and evaluate on unseen data
b)
To eliminate the need for any model
c)
To make all ratings identical
d)
To hide the item set from the algorithm
e)
To ensure no user is included in training
15.
Which metric family is listed for ranking evaluation?
a)
Precision, Recall, and F-score
b)
CPU utilization, memory, and disk
c)
Latency, throughput, and jitter
d)
Accuracy, AUC, and log loss only
e)
Silhouette score and inertia
16.
What does MAE measure in rating prediction evaluation?
a)
Average absolute difference between predicted and true ratings
b)
Average squared difference between predicted and true ratings
c)
Number of users in the dataset
d)
Number of items in the catalog
e)
Probability of the next click
17.
NMAE normalizes error using what information?
a)
The range between a user's maximum and minimum rating values
b)
The number of users in the test set
c)
The number of items rated by a user
d)
The number of latent factors K
e)
The learning rate
18.
RMSE is an evaluation metric based on what kind of error aggregation?
a)
Squaring errors before averaging and taking a root
b)
Averaging absolute errors only
c)
Counting only correct top-1 predictions
d)
Maximizing pageviews directly
e)
Using only binary labels
19.
In A/B testing for a recommender system, what is compared?
a)
Original system versus recommender system using user metrics
b)
Two different database engines for storage
c)
Two different web browsers for rendering
d)
Two different operating systems for deployment
e)
Two different network protocols for routing
20.
kNN collaborative filtering is based on what data structure?
a)
A user-item interaction matrix
b)
A decision tree of user profiles
c)
A convolutional image feature map
d)
A web page hyperlink graph only
e)
A dictionary of topic models only
21.
In user-based recommendation, neighbors are selected based on what?
a)
Similarity between users
b)
Similarity between item descriptions only
c)
Random sampling of users
d)
Alphabetical order of usernames
e)
Physical distance between users
22.
In computing user similarity, what is C?
a)
The set of items that both users rated
b)
The set of users that rated an item
c)
The set of all items in the catalog
d)
The set of timestamps in the dataset
e)
The set of latent factors
23.
In kNN rating prediction, what does V represent?
a)
The top k similar users to a target user
b)
The list of all items a user has never seen
c)
The set of all products in the catalog
d)
The set of all users in the system
e)
The validation set only
24.
What is a possible issue in neighbor-based rating prediction without constraints?
a)
A predicted rating may fall outside the normal rating scale
b)
Predictions always equal the global average
c)
The model cannot use similarity at all
d)
The test set becomes the training set
e)
The number of users becomes zero
25.
One disadvantage of user-based kNN collaborative filtering is what?
a)
It requires computation over the entire set of users
b)
It cannot use any interaction data
c)
It always needs item descriptions
d)
It only works for session-based data
e)
It produces only binary outputs
26.
Item-based recommendation is especially suitable when which condition holds?
a)
The number of products is much smaller than the number of users
b)
The number of users is always smaller than the number of products
c)
There are no products in the catalog
d)
Users never change over time
e)
All users rate every item
27.
A benefit of item-based recommendation is what?
a)
Item-item similarity can be calculated in advance
b)
User-user similarity must be recalculated for every click
c)
It removes the need for any evaluation
d)
It requires no data about items or users
e)
It prevents new items from being recommended
28.
In item similarity computation, what does U represent?
a)
Users who rated both items
b)
Items that both users rated
c)
All users in the system
d)
All items in the system
e)
Only new users
29.
In item-based rating prediction, what does J represent?
a)
The top k similar items to a target item
b)
The set of all users
c)
The set of all timestamps
d)
The set of all latent factors
e)
The set of all sessions
30.
In item-based collaborative filtering, predictions rely on what user data?
a)
The user's ratings of similar items
b)
Only the item title text
c)
Only the item price
d)
Only the user location
e)
Only the global average rating
31.
In matrix factorization, users and items are represented as what?
a)
Vectors in a latent aspect space
b)
Strings of product descriptions
c)
Decision rules in a tree
d)
A single global constant
e)
A list of web links
32.
In SVD-style factorization, what does K represent?
a)
The number of latent aspects
b)
The number of users
c)
The number of items
d)
The number of sessions
e)
The number of test examples
33.
In matrix factorization, a predicted rating is commonly computed using what operation?
a)
Dot product of a user vector and an item vector
b)
Counting common words in reviews
c)
Sorting items alphabetically
d)
Computing a web page PageRank score
e)
Measuring network latency
34.
In matrix factorization learning, which technique is mentioned to minimize error?
a)
Gradient descent
b)
Breadth-first search
c)
Apriori association mining
d)
Decision stump boosting
e)
Hash table lookup
35.
In gradient descent, what does the gradient indicate?
a)
How to change parameters to reduce error
b)
How to label items into classes
c)
How to compute exact similarity without data
d)
How to remove all noise from ratings
e)
How to avoid splitting data into train and test
36.
In model updates, what parameter controls the step size in gradient descent?
a)
Learning rate
b)
Number of users
c)
Number of items
d)
Precision
e)
Recall
37.
In model updates, what is the purpose of the regularization parameter lambda?
a)
Reduce overfitting by penalizing large parameter values
b)
Increase the rating scale from 1-5 to 1-10
c)
Replace gradient descent with random search
d)
Convert explicit ratings into clicks
e)
Guarantee perfect predictions
38.
Why does neural collaborative filtering use deep neural networks?
a)
To learn complex user-item interactions from basic to abstract
b)
To eliminate the need for any loss function
c)
To make all user vectors identical
d)
To avoid using embeddings
e)
To rely only on global averages
39.
One limitation of matrix factorization mentioned is that it preserves what?
a)
Similarity of user vectors
b)
Exact transaction timestamps
c)
Web page layouts
d)
Item images
e)
Database schemas
40.
In the limitations discussion, which similarity measure is assumed for users?
a)
Jaccard metric
b)
Euclidean distance on images
c)
Cosine similarity on text only
d)
Edit distance on usernames
e)
Hamming distance on URLs
41.
In the neural collaborative filtering architecture, what is the role of the embedding layer?
a)
Map one-hot inputs to dense latent vectors
b)
Compute test set size
c)
Split data into train and test
d)
Compute item prices
e)
Draw charts of traffic
42.
In neural collaborative filtering, how is a user represented at the input layer?
a)
A one-hot vector of length M
b)
A K-dimensional dense vector directly
c)
A list of item titles
d)
A sequence of web pages visited
e)
A binary matrix of sessions
43.
In the embedding layer description, what does an M by K weight matrix represent?
a)
User embeddings
b)
Item embeddings
c)
The test set labels
d)
The ranking metrics
e)
The train-test split
44.
What is the purpose of using a multi-layer perceptron in neural collaborative filtering?
a)
Learn nonlinear interactions between users and items
b)
Compute the global average rating
c)
Store raw transaction logs
d)
Precompute item-item similarities
e)
Replace embeddings with one-hot outputs
45.
For implicit feedback in neural collaborative filtering, what loss function is used?
a)
Cross-entropy loss
b)
Mean absolute error
c)
Root mean square error
d)
Hinge loss on images
e)
K-means objective
46.
Session-based recommendation is useful when what is difficult?
a)
Identify users and collect reviews
b)
Store item descriptions
c)
Compute cosine similarity
d)
Collect item prices
e)
Build a query interface
47.
In session-based recommendation, the model input is typically what?
a)
A sequence of previously interacted items
b)
A single user profile vector
c)
A list of explicit star ratings only
d)
A set of item images only
e)
A table of product prices only
48.
What does implicit feedback typically indicate?
a)
Whether a user interacted with an item
b)
A user's detailed written review
c)
An item's color and shape
d)
A user's demographic attributes
e)
An exact rating on a 1-5 scale
49.
In an RNN-based session model, what is the model asked to predict?
a)
The next item in the sequence
b)
The city where the user lives
c)
The total number of users
d)
The exact time a product ships
e)
The label of a document topic
50.
In the session model input layer, how is an item at time t represented?
a)
A one-hot vector of length V
b)
A float timestamp only
c)
A text description only
d)
A latent vector of length M
e)
A user-user similarity list
51.
In the session model embedding layer, one-hot vectors are transformed into what?
a)
K-dimensional representations
b)
M-dimensional one-hot vectors
c)
A list of item categories only
d)
A set of similarity scores only
e)
A binary rating matrix
52.
What is the purpose of the recurrent layer in a session-based recommender?
a)
Store historical information across time steps
b)
Compute item-item similarity in advance
c)
Remove cold-start for new items
d)
Convert explicit ratings to binary labels
e)
Split users into A/B groups
53.
What is the purpose of the MLP layer after the recurrent layer in the session model?
a)
Generate predictions over candidate items
b)
Compute rmax and rmin for normalization
c)
Store raw click logs without processing
d)
Compute gradients for user-based kNN
e)
Replace the embedding layer
54.
In pair-wise rank loss, what is the key assumption about a selected item?
a)
It should be ranked higher than unselected items
b)
It should always have a perfect explicit rating
c)
It should be removed from the training data
d)
It should be identical to the global average
e)
It should never appear in a session
55.
Which recommender approach does not require identifying individual users?
a)
Session-based recommendation
b)
User-based kNN collaborative filtering
c)
Matrix factorization with explicit ratings
d)
Content-based filtering using user profiles
e)
A/B testing
56.
If a user cannot express what they want with a query, which approach is more appropriate?
a)
Recommender system
b)
Information retrieval only
c)
Operating system scheduling
d)
Database sharding
e)
Web caching
57.
If transactions are very small compared to the number of users and products, what property of the interaction data is most likely?
a)
Sparsity in the user-item matrix
b)
Perfect coverage of all user-item pairs
c)
No need for evaluation metrics
d)
Guaranteed cold-start avoidance
e)
All users rate all items
58.
Cold-start is caused by little information about new users or new items. Which signal is most directly usable in a content-based method to reduce cold-start?
a)
Item features and user profile information
b)
User-user similarity from co-rated items
c)
Item-item similarity from historical ratings only
d)
A/B testing traffic split ratios
e)
Gradient descent learning rate
59.
If users and products change seasonally, which modeling consideration from the lecture becomes especially important?
a)
Time effects and periodic updates
b)
Ignoring timestamps completely
c)
Replacing all personalization with a global average
d)
Using only random recommendations
e)
Avoiding any train-test split
60.
When feedback is represented as 0 or 1 for interaction, what kind of prediction task does this resemble?
a)
Binary classification on implicit feedback
b)
Multiclass topic classification
c)
Image segmentation
d)
Graph shortest path computation
e)
Database normalization
61.
Why is RMSE typically more sensitive to large errors than MAE?
a)
RMSE squares errors, amplifying large deviations
b)
RMSE ignores all errors above a threshold
c)
RMSE counts only ranking mistakes
d)
RMSE uses only rmax and rmin
e)
RMSE does not use the test set
62.
What is a key benefit of using NMAE instead of MAE when different users use rating scales differently?
a)
It makes errors more comparable by normalizing by rating range
b)
It guarantees higher CTR in A/B testing
c)
It removes the need for a test set
d)
It automatically learns embeddings without training
e)
It prevents cold-start for new items
63.
Why might a system still need A/B testing even if offline MAE or RMSE is strong?
a)
Online user behavior metrics like CTR may not align with offline accuracy
b)
A/B testing replaces the need for any model training
c)
Offline metrics cannot be computed on ratings
d)
A/B testing is required to compute similarities
e)
A/B testing is needed only for content-based methods
64.
In user-based kNN, what operational cost is implied when a user makes new transactions frequently?
a)
User vectors need regular updates, increasing computation
b)
Item vectors cannot be updated at all
c)
The interaction matrix becomes dense automatically
d)
Regularization is no longer needed
e)
The model switches to session-based mode
65.
For a platform where the number of products is much smaller than the number of users, which neighbor-based approach is usually more practical and why?
a)
Item-based, because item vectors change less and similarities can be precomputed
b)
User-based, because it avoids computing any similarities
c)
Content-based, because it requires no item information
d)
Session-based, because it needs explicit ratings only
e)
Hybrid, because it prevents any sparsity
66.
If item-item similarities can be computed in advance, what effect does this tend to have on online recommendation latency?
a)
It reduces online computation and can speed up serving
b)
It increases the need to recompute all user vectors per request
c)
It makes gradient descent slower at training time
d)
It forces the model to use only explicit ratings
e)
It prevents any ranking evaluation
67.
User similarity is computed on the set of co-rated items. What issue arises if two users have very few co-rated items?
a)
Similarity estimates can be noisy and unreliable
b)
Similarity becomes exactly 1 for all user pairs
c)
The system no longer needs a test set
d)
The rating scale changes from 1-5 to 0-1 automatically
e)
Item-based recommendation becomes impossible
68.
Why can mean-centering user ratings help in neighbor-based rating prediction?
a)
It accounts for different user rating baselines when combining neighbor opinions
b)
It removes the need for any similarity computation
c)
It guarantees that predictions stay within the rating scale
d)
It converts explicit ratings into implicit feedback
e)
It replaces MSE with cross-entropy
69.
If a kNN model produces predicted ratings outside the allowed range, what simple post-processing step can keep outputs valid?
a)
Clamp predictions to the minimum and maximum rating values
b)
Increase the number of users in the system
c)
Remove the train-test split
d)
Replace similarity with random scores
e)
Convert all ratings to one-hot vectors
70.
Matrix factorization represents users and items in K latent aspects. What trade-off often comes with choosing a larger K?
a)
More expressive power but higher risk of overfitting and more computation
b)
Lower model capacity but higher bias
c)
Elimination of the need for regularization
d)
Guarantee that similarity is measured by Jaccard
e)
Removal of the need for any embeddings
71.
In gradient descent learning, what can happen if the learning rate is set too high?
a)
Training can become unstable and fail to converge
b)
The model automatically finds the global optimum in one step
c)
Regularization becomes unnecessary
d)
The test set becomes part of the training set by design
e)
Precision and recall cannot be computed
72.
What is the most direct purpose of regularization in matrix factorization learning?
a)
Reduce overfitting by discouraging overly large parameter values
b)
Increase the number of users and items in the dataset
c)
Turn explicit ratings into binary interactions
d)
Make the interaction matrix fully observed
e)
Guarantee higher pageviews in A/B testing
73.
The lecture notes that MF is a basic form of a neural network. What capability does deep neural collaborative filtering add beyond MF?
a)
Model nonlinear user-item interactions with multiple layers
b)
Eliminate the need for embeddings
c)
Remove the need for any loss function
d)
Force all users to share the same vector
e)
Prevent any use of implicit feedback
74.
NCF uses one-hot vectors for users and items, which can be very high dimensional. Which layer is designed to convert these into compact representations?
a)
Embedding layer
b)
A/B testing layer
c)
Train-test split layer
d)
Precision layer
e)
Recall layer
75.
Why is cross-entropy a natural loss for NCF with implicit feedback?
a)
The target is 1 or 0, so the task is binary prediction of interaction probability
b)
It directly minimizes RMSE on star ratings
c)
It requires only rmax and rmin
d)
It is used only for user-based kNN
e)
It avoids the need for negative samples
76.
Session-based recommendation uses transaction order for training. What does this allow the system to avoid collecting?
a)
Persistent user identification
b)
Any information about items
c)
Any interaction sequences
d)
Any model parameters
e)
Any evaluation metrics
77.
In the session-based problem definition, why is the output typically a ranked list of products instead of a single product?
a)
Because the goal is to return items with the highest next-item probability, often as top-N suggestions
b)
Because sessions cannot be represented as sequences
c)
Because ranking metrics cannot be computed otherwise
d)
Because RMSE requires a ranked list
e)
Because user profiles must be included in the output
78.
Why are LSTM or GRU mentioned as advanced options for the recurrent layer in session-based models?
a)
They help capture longer-term dependencies in sequences
b)
They remove the need for embeddings
c)
They guarantee perfect precision and recall
d)
They convert explicit ratings into one-hot vectors
e)
They eliminate the need for negative sampling
79.
In the session-based architecture, the final MLP is described as having V neurons. What does V correspond to?
a)
The number of distinct products/items
b)
The number of users
c)
The number of latent factors
d)
The number of train examples
e)
The number of evaluation metrics
80.
Pair-wise rank loss uses negative samples during training. Why are negative samples needed in implicit-feedback ranking?
a)
To teach the model that the observed item should be ranked above unobserved alternatives
b)
To increase the number of explicit ratings per user
c)
To compute MAE without a test set
d)
To avoid computing item embeddings
e)
To force the interaction matrix to be dense
81.
When would hybrid recommendation methods be especially useful, based on the lecture topics?
a)
When combining content signals with collaborative signals can reduce cold-start and sparsity issues
b)
When you want to avoid using any user or item data
c)
When you only need to sort items alphabetically
d)
When you cannot store item features at all
e)
When evaluation metrics are not required
82.
Collaborative filtering depends on similarity learned from history. What is a key risk if user consumption habits shift seasonally?
a)
Similarities can become stale, so recommendations may degrade unless updated over time
b)
The system no longer needs train and test sets
c)
The interaction matrix becomes fully observed
d)
Precision and recall become identical by definition
e)
Cross-entropy becomes invalid for implicit feedback
83.
For a news site where users are often anonymous, which approach from the lecture is the best fit and why?
a)
Session-based, because it does not require user identification and uses click sequences
b)
User-based kNN, because it needs explicit star ratings
c)
Matrix factorization, because it needs stable user IDs
d)
Content-based only, because it uses community data
e)
A/B testing, because it is itself a recommendation method
84.
If offline RMSE improves but CTR drops in an A/B test, what is the most reasonable interpretation?
a)
Offline rating accuracy may not match the online objective, so evaluation and objectives may be misaligned
b)
The train set is always larger than the test set
c)
Precision and recall cannot be computed
d)
The item set I was defined incorrectly
e)
The model has no parameters to tune
85.
Ranking metrics like precision and recall are most appropriate when the main goal is what?
a)
Recommend top-N items rather than predict exact rating values
b)
Estimate rmax and rmin for each user
c)
Compute derivatives for gradient descent
d)
Precompute item vectors without data
e)
Remove seasonal effects completely
86.
A platform has tens of millions of users and far fewer items, and it needs fast online recommendations. Which neighbor-based strategy best reduces online computation while staying consistent with the lecture points?
a)
Item-based collaborative filtering with precomputed item-item similarity
b)
User-based collaborative filtering that recomputes similarities for every request
c)
Content-based filtering that ignores item features
d)
A/B testing as the primary recommendation algorithm
e)
Session-based recommendation that requires persistent user IDs
87.
User-based kNN computes similarity using co-rated items. In an extremely sparse interaction matrix, what failure mode is most likely?
a)
Neighbors may be chosen based on noisy similarity from too few co-rated items, leading to unstable predictions
b)
All user-user similarities become exactly zero, making predictions always perfect
c)
The model automatically becomes a matrix factorization model
d)
The test set becomes unnecessary because sparsity improves evaluation
e)
Item-item similarities can no longer be computed in advance
88.
A matrix factorization model trained with gradient descent shows lower training error but higher test error. Which adjustment is most directly aimed at this issue in the lecture material?
a)
Increase the regularization parameter lambda
b)
Increase the learning rate far beyond its current value
c)
Remove the train-test split
d)
Switch the rating scale from 1-5 to 0-1 without changing the loss
e)
Use only the global average as the prediction
89.
Neural collaborative filtering represents users and items as one-hot vectors of size M and N. For very large M and N, which component is most likely to dominate memory usage?
a)
The embedding weight matrices of size M by K and N by K
b)
The cross-entropy loss computation
c)
The output probability scalar
d)
The precision and recall calculations
e)
The train-test split procedure
90.
A team evaluates a model with MAE on explicit ratings but deploys it for click prediction using implicit feedback. What is the most serious issue with this setup?
a)
The training objective and evaluation signal may not match the deployed feedback type, hurting real performance
b)
MAE cannot be computed on any dataset
c)
Implicit feedback cannot be used with neural networks
d)
Clicks always provide a 1-5 scale automatically
e)
A/B testing becomes impossible to run
91.
In implicit feedback data, an unobserved item is often treated as 0. Why can this assumption be risky for learning and evaluation?
a)
Unobserved does not necessarily mean disliked, so treating it as negative can introduce false negatives
b)
Unobserved items always have the maximum rating
c)
Unobserved items make the interaction matrix fully dense
d)
Unobserved items remove the need for any negative samples
e)
Unobserved items guarantee perfect precision
92.
Two users have the same MAE, but one user uses only a narrow personal rating range while the other uses a wide range. Which metric from the lecture better supports fair comparison across these users?
a)
NMAE, because it normalizes by the user's rating range
b)
RMSE, because it ignores rating range
c)
Precision, because it measures exact rating errors
d)
Recall, because it measures squared error
e)
F-score, because it is a regression metric
93.
A user-based kNN predictor outputs values outside the rating scale. What does this most strongly indicate about the prediction formula behavior?
a)
It is using an unbounded weighted combination that can exceed rating limits unless normalized or clamped
b)
It is guaranteed to output the global average rating
c)
It has converted explicit ratings into one-hot vectors
d)
It has removed the similarity computation step entirely
e)
It has replaced MAE with cross-entropy
94.
The lecture highlights that preferences and consumption habits can change seasonally. For a domain with strong seasonality, which enhancement is most critical to consider and why?
a)
Time effects, because the relevance of past interactions can change over time
b)
Only global averages, because personalization hurts seasonality
c)
Only item-item similarity, because it never changes
d)
Only train-test splitting, because it replaces temporal modeling
e)
Only binary scaling, because it removes timestamps
95.
A content-based system often recommends items similar to what the user already liked. Which method category from the lecture is most suitable to broaden recommendations using peer behavior?
a)
Collaborative filtering or a hybrid method
b)
Content-based only, without any additional signals
c)
A/B testing, as a recommendation method
d)
Train-test splitting, as a recommendation method
e)
Gradient descent, as a recommendation method
96.
Item-based recommendation assumes item vectors change less frequently and similarities can be precomputed. What operational problem arises if the item catalog and interactions change rapidly in real time?
a)
Precomputed item-item similarities can become stale and require frequent recomputation
b)
The interaction matrix automatically becomes dense, removing the need for similarity
c)
User identification becomes impossible by definition
d)
Cross-entropy loss can no longer be used
e)
The number of users must be smaller than the number of items
97.
User-based kNN requires updating user vectors with new transactions and computing over many users. Under a strict real-time requirement, what is the dominant scalability bottleneck?
a)
Computing similarities across a very large user set and updating neighbors frequently
b)
Computing rmax and rmin for each user once
c)
Computing item-item similarity once per year
d)
Computing precision and recall on the training set only
e)
Computing the agenda of the lecture
98.
A session-based RNN model represents each item as a V-dimensional one-hot vector. If V is extremely large, what is the most direct computational consequence mentioned by the architecture design?
a)
The model needs large embedding-related parameters, since one-hot inputs are high dimensional
b)
The model no longer needs an embedding layer
c)
The model outputs only binary labels without probabilities
d)
The model cannot use recurrent layers
e)
The model becomes a content-based recommender
99.
Pair-wise rank loss assumes selected items should outrank other items. With only positive interactions observed, what key ambiguity remains in implicit feedback data?
a)
It is unclear whether unobserved items are truly negative or simply not seen
b)
All unobserved items are guaranteed to be irrelevant
c)
All users have identical preferences
d)
All items have the same embedding
e)
All sessions have the same length
100.
You train NCF with cross-entropy on implicit feedback where non-interactions vastly outnumber interactions. Which evaluation approach from the lecture best reflects recommendation quality in this setting?
a)
Ranking metrics such as precision and recall for top-N recommendations
b)
Only MAE, because it measures click probability directly
c)
Only RMSE, because it is designed for binary labels
d)
Only NMAE, because it is a classification loss
e)
Only the global average rating baseline
101.
When interaction data is sparse, what combination of modeling choices is most conservative to reduce overfitting in matrix factorization?
a)
Use a smaller K and stronger regularization lambda
b)
Use a larger K and remove regularization
c)
Use a larger learning rate and remove the test set
d)
Replace user and item vectors with one-hot outputs
e)
Use only A/B testing without a model
102.
In an A/B test, users are split between the original system and the recommender system. What validity risk occurs if the treatment model is retrained using treatment user behavior during the experiment without proper controls?
a)
Contamination can confound the comparison, because the treatment changes over the test period in a way tied to the split
b)
It guarantees that CTR will always increase
c)
It makes MAE impossible to compute
d)
It removes the need for any evaluation metrics
e)
It forces the model to become session-based
103.
Hybrid methods combine content-based and collaborative filtering. For new items with no interactions yet, which component provides the primary recommendation signal and why?
a)
Content-based, because item features can be used even without historical interactions
b)
User-based kNN, because it needs co-rated items for the new item
c)
Item-based kNN, because item similarity requires many past ratings
d)
A/B testing, because it generates item features automatically
e)
RMSE, because it is an item representation method
104.
The lecture notes limitations of MF related to fixed similarity assumptions. How does deep NCF most directly address this limitation?
a)
By learning a flexible nonlinear interaction function through multilayer networks
b)
By removing user and item representations entirely
c)
By replacing embeddings with global averages only
d)
By requiring explicit queries like information retrieval
e)
By eliminating the need for any loss function
105.
Consider an application focused on predicting the next item in a short anonymous browsing session. Which combination of lecture concepts best fits this goal?
a)
Session-based recommendation with an RNN-style sequence model and implicit feedback training
b)
User-based kNN with long-term user profiles and explicit ratings only
c)
Matrix factorization trained only on explicit 1-5 ratings with no sequence information
d)
Content-based filtering that ignores transaction order
e)
A/B testing as the prediction model
100 %
