Description: Recommendation engines have boomed in the era of social media. These panelists are experts at collaborative filtering systems. Citing Digg, Amazon, and Netflix as examples, they will have a high level discussion about the evolution of recommendation engines and how each approach is different.
Speakers: Anton Kast (Lead Scientist, Digg), Scott Brave (CTO, Baynote Inc), David Maher Roberts (CEO, The Filter), Jon Sanders (Dir Recommendation Systems, Netflix), Erik Frey (Research Engineer, Last.fm)
Anton Kast
-Collaborative filtering has deep academic roots , Xerox, UMN, MIT
-Definition: Combines the input from many different people to filter the information better than would otherwise be possible
-Technique is everywhere from spam filter to comment moderation to tagging in facebook ads to pagerank
When it is personalized, it is "recommendation" ... Amazon, Netflix, behavioral ad targeting
-How DIGG works: anyone can submit, vote on any story; can view most popular stories
Problems: the sparsity problem (not enough) early rater problem, gray sheep problem (what's not popular but a small group likes...how do you serve?) user opposition
Eric Frey
-Last.fm is social music website with 25m ussers
-2 sets of relationships: songs and people; people and people
-Different types of reco's provide different context
--Lean forward reco's: user is engaged in site, want to discover new music
--Lean Back reco's: continuity is important
-Focus on mining better data rather than developing complex algorithms ... Data is the most important ingredient (at last.fm it's social tags and listening data)
The recipe: can increase reco's by how you model the user to build their attention profile and figure out what they are interested in
Scott Brave
-Reco's offered as a service to other websites
-Use java to capture info on site, how they got there and what they do on site
-Data feeds into affinity engine and they look for value
-Social search - reordering of results based on what people are using
David Maher Roberts
-Filter is a digital entertainment content recommendation and relevance engine, focus on video and audio, delivers as white label serviuce, 10m reco's a day
-Visualize data via a "taste cloud"
-Created an entertainment dna
Jon Sanders
-60% or purchases happen via reco's
-10 years ago netflix was a stand alone site, first introduced rating widget, then looked at how to make reco's more credible, then asked about interest, then asked other people who write reviews, etc., finally started explaining why we were recommending something or showing user a genre
-Million dollar competition going on
-More data is always goo
-CF is a component of personalization
0 comments:
Post a Comment