How do recommender systems work?

A simple introduction

Yanir Seroussi

yanirseroussi.com | @yanirseroussi | linkedin.com/in/yanirseroussi

Note: I also published a more thorough article on recommender systems.

No magic

GIGO: Any data-driven system can only be as good as the data it's fed

Fancy mathematical models often try to emulate simple intuition

Classic scenario

Question: Given a person, what movies would they like?

Answer: depends on the data

No data

We know nothing about movies and nothing about the person

Best guess: random movies

Partial rating* data

We know what others rated, but nothing about the person

Best guess: popular movies

Richer rating data

We have ratings by the person and by other people


Collaborative filtering

Find people with similar taste to the person and recommend based on their taste

Find movies that are rated similarly to the movies liked by the person

Richer rating data + movie metadata

Content-based recommendation: find similar movies to liked movies based on genre, cast, etc.

Can be combined with collaborative filtering

And beyond...

Social data

Context

Temporal data

Price

Other considerations

Diversity of recommendation list

Serendipity versus trust in the system

Short-term gains versus lifetime value

Relevance to the user versus revenue

User interface and experience

Bottom line: build-measure-learn