Recommender systems quiz

Question 1

In collaborative filtering, what fundamental problem does the 'cold start' refer to, and name its two main variants.

Accepted Answer

Insufficient data for new users or new items, with user and item cold-start variants

Answer

The latency of serving recommendations at startup time

Answer

Overfitting that occurs early in model training

Answer

The bias toward popular items in the catalog

Question 2

Why is matrix factorization (e.g., SVD-style latent factor models) preferred over memory-based neighborhood methods for large, sparse rating matrices?

Accepted Answer

It generalizes across sparse data via low-rank latent factors and scales better

Answer

It requires no hyperparameter tuning

Answer

It guarantees exact reconstruction of the rating matrix

Answer

It eliminates the need for any regularization

Question 3

In implicit-feedback recommendation, why can't you treat unobserved interactions simply as negative labels, and how does the Weighted ALS (Hu, Koren, Volinsky) approach address this?

Accepted Answer

Unobserved pairs are ambiguous, so WALS adds confidence weights and treats them as weak negatives

Answer

Unobserved entries are dropped entirely from the loss

Answer

Missing entries are imputed with the global mean before training

Answer

Negative sampling is mandatory and WALS forbids it

Question 4

What is the key conceptual difference between BPR (Bayesian Personalized Ranking) loss and pointwise regression loss for recommendation?

Accepted Answer

BPR optimizes pairwise ordering of positive vs sampled negative items rather than absolute scores

Answer

BPR predicts absolute ratings while pointwise predicts ranks

Answer

BPR can only be used with explicit ratings

Answer

They are mathematically identical losses

Question 5

In a two-tower retrieval model, why are the user and item towers kept separate with a dot-product (or cosine) interaction at the top?

Accepted Answer

To precompute and ANN-index item embeddings, enabling sublinear retrieval

Answer

To allow gradient flow between towers during inference

Answer

Because dot products are more accurate than neural networks

Answer

To avoid having to train the item tower

Question 6

Explain the difference between the retrieval (candidate generation) and ranking stages in an industrial multi-stage recommender.

Accepted Answer

Retrieval cheaply maximizes recall over many items; ranking precisely orders a small candidate set

Answer

Retrieval ranks final order; ranking just filters spam

Answer

They are two names for the same model

Answer

Ranking runs before retrieval in the pipeline

Question 7

What is the 'feedback loop' / popularity bias problem in deployed recommenders, and name one mitigation.

Accepted Answer

Logged exposure amplifies popular items in a self-reinforcing loop; mitigate with IPS or exploration

Answer

Models forget popular items over time; fix with more epochs

Answer

It only affects content-based systems and needs no fix

Answer

It is solved automatically by larger embedding dimensions

Question 8

In learning-to-rank for recommendation, what does NDCG@k measure and why is it preferred over precision@k?

Accepted Answer

NDCG is position-discounted and graded, rewarding relevant items ranked higher, unlike binary precision@k

Answer

NDCG ignores position; precision@k accounts for it

Answer

NDCG only works for implicit feedback

Answer

They are equivalent for k>10

Question 9

Why is offline AUC or NDCG often a poor predictor of online A/B test lift in recommenders?

Accepted Answer

Logged data is biased by the old policy and labels are missing-not-at-random, breaking the offline-online correlation

Answer

Offline metrics are always lower than online metrics by a constant

Answer

Online tests measure latency only

Answer

AUC is uncomputable on recommendation data

Question 10

What problem does negative sampling solve when training large-scale retrieval models with softmax over the full item catalog?

Accepted Answer

It makes the full softmax tractable by approximating the partition function with sampled negatives

Answer

It removes the need for embeddings

Answer

It guarantees zero gradient variance

Answer

It only applies to explicit-rating models

Question 11

In sampled softmax / in-batch negatives, why is a logQ correction applied, and what is being corrected?

Accepted Answer

To debias logits since popular items are oversampled as negatives, correcting toward the true softmax

Answer

To increase the learning rate adaptively

Answer

To normalize embeddings to unit norm

Answer

To regularize the item tower weights

Question 12

What is the purpose of the FM (Factorization Machine) model, and what does it add over linear models?

Accepted Answer

FM factorizes pairwise interaction weights so it generalizes to sparse/unseen feature combinations

Answer

FM only models linear effects faster

Answer

FM requires dense features to work

Answer

FM replaces embeddings with one-hot encodings

Question 13

In the Wide & Deep architecture, what role does each component play?

Accepted Answer

Wide memorizes specific cross-features; deep generalizes via embeddings

Answer

Wide does generalization, deep does memorization

Answer

Both components are identical MLPs

Answer

Wide handles images, deep handles text only

Question 14

What is the core idea behind sequential recommendation models like SASRec, and why use self-attention?

Accepted Answer

Self-attention adaptively weights relevant past items to predict the next item in a sequence

Answer

It clusters users by demographics

Answer

It removes the need for item embeddings

Answer

It only models the most recent item

Question 15

How does BERT4Rec differ from SASRec in its training objective?

Accepted Answer

BERT4Rec uses a bidirectional masked (Cloze) objective; SASRec uses causal next-item prediction

Answer

BERT4Rec is causal; SASRec is bidirectional

Answer

Both use identical objectives but different optimizers

Answer

BERT4Rec does not use attention

Question 16

What is the exploration-exploitation tradeoff in recommendation, and how do contextual bandits address it?

Accepted Answer

Contextual bandits balance trying uncertain items vs serving best-known ones using context and uncertainty

Answer

Bandits eliminate the need to ever explore

Answer

The tradeoff only matters offline

Answer

Exploration always reduces long-run reward

Question 17

Why is diversity an important objective alongside accuracy, and what is one method to inject it at ranking time?

Accepted Answer

Redundant lists hurt satisfaction; MMR or DPP trade relevance against intra-list similarity

Answer

Diversity is irrelevant once accuracy is high

Answer

Diversity is only achievable by random shuffling

Answer

It is enforced by increasing embedding dimension

Question 18

What is inverse propensity scoring (IPS) in counterfactual recommendation evaluation, and what is its main weakness?

Accepted Answer

IPS gives unbiased off-policy estimates but suffers high variance for small propensities

Answer

IPS is biased but low variance by design

Answer

IPS requires online traffic to compute

Answer

IPS only works for explicit ratings

Question 19

In graph-based recommendation (e.g., LightGCN), what simplification did LightGCN make over standard GCNs and why?

Accepted Answer

It removed feature transformations and nonlinearities, keeping only neighborhood aggregation

Answer

It added more nonlinear layers for capacity

Answer

It replaced the graph with a sequence model

Answer

It uses only first-order neighbors and no propagation

Question 20

What is the 'filter bubble' concern in recommenders, and how does it relate to serendipity?

Accepted Answer

Personalization narrows exposure to homogeneous content; serendipity (relevant + unexpected) counters it

Answer

Filter bubbles increase content diversity automatically

Answer

Serendipity means recommending only popular items

Answer

Filter bubbles are a hardware caching issue

Question 21

When using embeddings for categorical IDs in a deep recommender, why is the hashing trick sometimes used, and what is its risk?

Accepted Answer

It bounds embedding memory and handles new IDs but risks hash collisions conflating distinct IDs

Answer

It increases accuracy by removing rare IDs

Answer

It eliminates the need for embeddings entirely

Answer

It only applies to numeric features

Question 22

What is position bias in click-through data, and how can a ranking model be debiased for it?

Accepted Answer

Higher ranks attract clicks regardless of relevance; debias via IPS examination model or a position feature

Answer

Position bias means lower items get more clicks

Answer

It is removed by sorting items alphabetically

Answer

It only affects retrieval, not ranking

Question 23

Why do many production recommenders calibrate predicted probabilities (e.g., for CTR), and what is one calibration method?

Accepted Answer

Predicted probabilities must match true rates for value-based decisions; use Platt/isotonic or sampling correction

Answer

Calibration changes the ranking order to improve AUC

Answer

Calibration is only needed for regression targets

Answer

It is unnecessary if AUC is high

Question 24

What is multi-task learning in ranking models (e.g., MMoE), and why use it for recommendation?

Accepted Answer

MMoE shares experts with per-task gates to optimize multiple objectives while limiting negative transfer

Answer

MMoE trains one task at a time sequentially

Answer

Multi-task learning forces all tasks to share identical heads

Answer

It only works when tasks are perfectly correlated

Question 25

How does L2 regularization in matrix factorization help, and what is its effect on latent factors?

Accepted Answer

It shrinks latent-factor norms, reducing overfitting and improving generalization on sparse data

Answer

It increases factor norms to add capacity

Answer

It removes the bias terms

Answer

It guarantees orthogonal factors

Question 26

What is the role of user and item bias terms in a matrix factorization rating model r_ui = mu + b_u + b_i + p_u . q_i?

Accepted Answer

They capture global, per-user, and per-item systematic effects so the dot product models true interaction

Answer

Biases add random noise for regularization

Answer

Biases replace the latent factors

Answer

They are only used for implicit feedback

Question 27

In approximate nearest neighbor (ANN) search for retrieval, what tradeoff does an HNSW index manage?

Accepted Answer

It trades recall against latency/memory via parameters like M and efSearch

Answer

HNSW guarantees exact nearest neighbors with no tradeoff

Answer

It only works with Euclidean distance

Answer

It eliminates the need for embeddings

Question 28

What is a common cause of training-serving skew in recommenders, and how is it mitigated?

Accepted Answer

Different feature computation paths/timing between train and serve; mitigate with a feature store and serve-time logging

Answer

It is caused by using too many GPUs

Answer

It only occurs with linear models

Answer

It is fixed by increasing batch size

Question 29

How can large language models (LLMs) be leveraged in modern recommender systems as of 2026?

Accepted Answer

LLMs aid semantic understanding, cold start, conversational/explainable rec, and re-ranking, complementing collaborative signal

Answer

LLMs fully replace collaborative filtering in all production systems

Answer

LLMs are only useful for image recommendation

Answer

LLMs remove the need for any candidate retrieval

Question 30

What is generative retrieval (e.g., TIGER with semantic IDs) and how does it differ from dual-encoder retrieval?

Accepted Answer

It autoregressively decodes item semantic-ID tokens, replacing ANN dot-product matching

Answer

It is just dual-encoder retrieval with a bigger index

Answer

It only works for explicit ratings

Answer

It requires no training data

Test yourself