Recommender systems
Nicole and Kirstin
What is a recommender system?
• A system tries to predict if a user will like an item.
• The system recommends to a user items that it thinks the user will like.
What would you want recommendations for?
How do you make recommendations?
• Your friend is looking for a new book to read.
• How can a system make recommendations?
Two types of information can inform recommendations.
Similar people have similar tastes.
People like related items.
We liked Harry Potter—you’ll love it!
I like historical fiction!
I like fantasy!
Similar people have similar tastes!
Peggy
Christine
Collaborative Filtering (CF)
People like related items!
Susan
Molly
Content-based filtering (CB)
Let’s do an example together
Peggy
Christine
Brandon
Peggy
Christine
Brandon
Ricky
Collaborative filtering looks at similar users
Peggy
Christine
Brandon
Ricky
Content-based looks at the items themselves
Cartoons
Which do you think works better?
Collaborative filtering: Similar people have similar tastes.
Content-based filtering: People like related items.
We liked Harry Potter—you’ll love it!
I like historical fiction!
I like fantasy!
Should you recommend this?
Christine
Ricky
Sashank
Should you recommend this?
Maria
Sue
Caitlin
What should you recommend?
Nancy
Hannah
Priya
Should you recommend this?
Nancy
Hannah
Priya
How are these items and people related?
Michelle
Katy
Jan
Alice
Anjali
Michelle
Katy
Jan
Alice
Anjali
Alex
Simone
Gina
Training Set
Sashank
Brandon
Maria
Christine
Susan
Jesse
Test Case 1: Michelle
Michelle
Test Case 2: Katy
Katy
Test Case 3: Jan
Jan
How can you improve these systems?
• What were some downfalls of each method?
– Collaborative filtering
• Hard to answer when new person hasn’t seen/read/bought anything that other people have
– Content-based filtering
• Hard to answer when someone doesn’t have a clear profile of interests
• Hybrid approaches combine elements of both collaborative and content-based filtering.
Netflix is mainly content-based
Netflix attempted to get better algorithms…
But the winner turned out to be too complicated to practically use!
Pandora is also content-based
“The Music Genome Project”
Amazon uses both
Where else could recommender systems be helpful?