Anton Kast listing common examples.
Spam filters, pagerank, tagging, comment moderation, help systems, facebook ads
Thumbs up/down on everything.
Recommendation: personalized. Custom search results, Amazon's recursive filtering of recommended products
Going over digg.com recommendation engine
Comparison of homepage vs. recommended stories. Correlated with users "like you"
Sparsity problem: submissions can grow faster than active diggers
Gray sheep problem: Small group of people very enthusiastic about unpopular views
Erik Frey from Last.fm takes over
Talking about discovering new music through last.fm
Relationships between songs and users and users and other users (people recommendations).
Algorithms and social channels.
Different algorithms for different contexts
Lean Forward recommendation: Engaged in site, interested in variety
Lean Back recommendation: not engaged, ambient, background
Data is the most important ingredient
Scrobbles + Social Tags + Love + Ban + Skip
25 billion scrobbles on Last.fm so far
Scrobbles are out of context and don't tell the whole story. Unknown artist found similar to Bethoven because his track was also in default sounds folder in win xp
More data points help form context and improve recommendations
Scott Brave of Baynote
Recommendation engine as a service on other websites
Javascript snippet to gain data
Tracks engagement
Affinity Engine
Re-orders search results based on user patterns
Too Many Results: technically correct results, but practically useless
50% give up after third result
Fulltext matches keywords, but doesn't yield useful results
Full-text processing, Meta-tagging Taxonomy, Folksonomy
Implicit Tagging: tracking keywords through user actions
David Mayer Roberts of The Filter
Entertainment content recommendation engine
bayesian collaborative filter model
UI Examples: Taste Cloud: visualization of implicit and explicit actions, infers other elements, creates feed
Great visualization: data points surrounding dart board like heart. Dearest items are shown closest to the center of the heart
Anonymized data to find relationships between verticals (music, video, etc)
Notion of an entertainment DNA
John Sanders of Netflix
60% of movie selections based on recommendation
First: The rating widget
Second: Take score and sort implied interest
Next: Built in k-nearest-neighbour algorithm to tie in movie likeness
Next: Added interest-based discovery + meta data connections (same actor/director)
Added genre ratings to explicitly collect interests
Found it valuable to explain why something was recommended based on past evidence.
Recommendations are domain specific, and need to be tailored to different forms of media
People want to drive and not be lead. Support actions vs. lead actions.
Lots of questions coming...
Popularity bias skews relationships
Book recommendation: Collective Intelligence from O'Reilly
Engines are built on similar priciples but are difficult to port due to the importance of context
Domain specific tweaks
Talking about global activities contributing to recommendations across platforms. ie. can netflix data influence last.fm recommendations
Question about filtering data which throws off the recommendation engine.
Discussing positive vs negative input: Thumbs up and thumbs down vs positive only input.
Thumbs up UI. Does thumbs up mean support for an item, or for the review?
AB Testing against single measurement
Geo-awareness is a good place to start if you don't know anything about someone
playground.last.fm has a list of the guiltiest scrobbles
and we're done.