Collaborative filtering uses input from many users to filter information and provide recommendations. It originated in research projects in the early 1990s that allowed users to rate messages and content. This technique is now widely used to provide personalized recommendations through systems like recommendation engines, spam filters, search engines, social tagging, and comment moderation. Recommender systems like Amazon and Netflix use collaborative filtering to provide personalized product and media recommendations to users.
2. Old Terms with Origins in Research
( user annotations, )
“Tapestry” email filter at Xerox PARC, 1992 search grammar
(readers rate messages,)
“GroupLens” usenet filter at UMN, 1994 correlates people
(social actual business, )
an
“Firefly” music and friend recommender at MIT, 1995 network elements
Research origins led to sophisticated implementations,
but the idea is simple.
Collaborative Filters: Recommendation Engines
3. The Idea of
Collaborative Filtering
Combine input from many different people to filter
information better than would otherwise be possible.
Collaborative Filters: Recommendation Engines
4. This Technique is Everywhere
Spam filters Help systems
Pagerank Click feedback in search ranking
Tagging Facebook ads
Comment moderation Thumbs on everything
Collaborative Filters: Recommendation Engines
5. When it’s Personalized
Call it ‘Recommendation’
Music, movie, book sales Customized search results
Behavioral ad targeting Google News
Amazon’s recursion: filtered recommendations!
Collaborative Filters: Recommendation Engines
6. How Digg Works
1. Anyone can submit a story
2. Anyone can vote on any story
3. Most popular recent stories win
(classic collaborative filtering)
4. If you sign up, you get personalized stories too
(recommendations)
Collaborative Filters: Recommendation Engines
7.
8.
9. Where we can Leave the Rails
The sparsity problem
Submissions can grow faster than active diggers
The early-rater problem
We have no way to jump-start the recommendation cycle
Gray sheep
We have smaller sub-communities with unpopular views
User opposition
A few times, we have simply come to loggerheads
10. Introductions
Anton Kast, Digg
Erik Frey, Last.fm
Scott Brave, Baynote
David Maher Roberts, TheFilter
Jon Sanders, Netflix