/
Flashcards from Upload
Save to my account
Sign up
Flashcards from Upload
Recommendation Systems Concepts
Study
1
Question
What is the primary purpose of approximate nearest neighbor (ANN) search?
Answer
To quickly find similar items in high-dimensional space.
2
Question
How do embedding representations facilitate feature engineering?
Answer
They convert qualitative features into numerical vectors for model input.
3
Question
What are static user features? Give an example.
Answer
Attributes that don't change frequently, e.g., device type.
4
Question
What are dynamic user features? Provide an example.
Answer
Attributes that change with recent activity, e.g., videos watched last hour.
5
Question
Give an example of a content feature.
Answer
Video length or upload date.
6
Question
How do interaction features differ from content features?
Answer
They describe the relationship between user and content, like previous interactions.
7
Question
What role do embeddings play for categorical features?
Answer
They capture similarities and generalize patterns for categories.
8
Question
Why are neural networks preferred over logistic regression in complex recommendation models?
Answer
They can model nonlinear, high-dimensional data and patterns.
9
Question
What is a two-tower model in recommendation systems?
Answer
Two separate networks process user and content features independently.
10
Question
Why is latency critical in recommendation scoring?
Answer
Recommendations must be computed in milliseconds for a seamless experience.
11
Question
What techniques are used to optimize model inference speed?
Answer
Model distillation, quantization, and pruning.
12
Question
How are raw scores from the prediction model transformed into the final ranked list?
Answer
Through weighting, diversity adjustments, and business rules.
13
Question
What is the purpose of business rules in content ranking?
Answer
They enforce policies and strategic priorities, overriding model predictions.
14
Question
How does user interaction feedback influence recommendation systems?
Answer
It updates user profiles, helping the system adapt over time.
15
Question
What is the main goal of diversity injection in recommendation feeds?
Answer
To prevent redundancy and maintain user engagement.
16
Question
How does YouTube optimize recommendations for long-form content?
Answer
By focusing on session duration and long-term retention.
17
Question
What key feature helps YouTube understand video content?
Answer
Structured metadata like titles, descriptions, and transcripts.
18
Question
What is a multi-task model in the context of recommendations?
Answer
A model predicting multiple engagement actions simultaneously.
19
Question
How does Instagram balance personal connections and content interests?
Answer
By weighing both the social graph and behavioral signals.
20
Question
What signals does Instagram Reels use for recommendations?
Answer
Replay rate, watch time, and audio features.
21
Question
What is unique about Spotify's recommendation system compared to social media?
Answer
It uses passive listening signals and content analysis of audio.
22
Question
How does Spotify analyze song features?
Answer
From spectrograms using convolutional neural networks.
23
Question
What does Spotify's two-tower neural network predict?
Answer
Likelihood of actions like listening, skipping, or replaying.
24
Question
What strategy does Spotify use to maintain playlist novelty?
Answer
Diversifying genres, tempos, and artists in recommendations.
25
Question
How does Netflix personalize content presentation?
Answer
Using contextual features and personalized artwork.
26
Question
What modeling approach does Netflix use for recommendation?
Answer
Contextual personalization and satisfaction proxies.
27
Question
Why does Netflix avoid optimizing solely for watch time?
Answer
To optimize long-term user satisfaction and retention.
28
Question
What is a key challenge unique to dating app recommendations?
Answer
Mutual interest prediction between two users.
29
Question
What problem does data sparsity cause in dating recommendations?
Answer
Limited explicit feedback and few interactions.
30
Question
How do dating apps estimate mutual compatibility?
Answer
Using user profile vectors and models predicting match likelihood.