강의

Background

Terminalogy

1. item

• entity for system to recommend
• youtube -> video
• google play store -> apps
• also known as document

2. Query

• information for system to use for recommendation
• user info + additional context

3. Embedding

• map discret set to vector space
• quries set, item set -> embedding space

common recommendation system architecture

1. candidate generation

• start from huge corpus -> generate smaller candidate subset
• can be multiple candidate -> each nomiate different candidate subset

2. scoring

• score/rank candidates -> select best subset that match with user
• model evaluate small item subset, so system use more precise model with additional queries

3. re-ranking

• remove explicity disliked
• boost fresher content score
• reflect additional constrains for final ranking
• ensure freshness, diversity, fairness