Design of recommender systemsPresentation Transcript
Design Strategies for Recommender Systems Rashmi Sinha www.uzanto.com Jan 2006, UIE Web App Summit
What are Recommender Systems?
Systems that attempt to predict items, e.g., movies, music, books, that a user may be interested in (given some information about the user's profile)
e.g., Amazon – people who liked this book also liked, Netflix recommendations
Systems that help people find information that will interest them, by facilitating social / conceptual connections or other means…
Designing different finding experiences
Some experiences guide user, others just point in a general direction
Desired experience depends on user task, time constraints, mood etc.
There’s more than one way to get from here to there…
User experience in search/browse interfaces
More controlled experience
Every movement (forward, making a turn) is a conscious choice
System should provide information at every step
If user takes wrong turn, go back a step or two / start again
Like driving a car…
User Experience with Recommender Systems
User has less control over specifics of interaction
System does not provide information about specifics of action
More of a “black box” model (some input from user, output from systems)
Like riding a roller coaster…
Recommender Systems Circa 2001
what movies you should watch… (Reel, RatingZone, Amazon)
what music you should listen to… (CDNow, Mubu, Gigabeat)
what websites you should visit… (Alexa)
what jokes you will like… (Jester)
where to go on vacation (TripleHop)
& who you should date… (Yenta)
I know what you will read next summer!
A technological proxy for a social process “ I think you would enjoy reading these books…” Friends / Family Ref: Flickr photostream: jefield Ref: Flickr-BlueAlgae Ref: Flickr-Lady_Strathconn What should I read next?
Interaction paradigm “ Books you might enjoy are…” Output: Input: Rate some books Ref Flickr photostreams: anjill154 & rossination What should I read next?
Meg & James: correlation = .52
How collaborative filtering algorithms work Recommendations For Meg Lets find a book for Meg!
Input : Motivating users to give input (to feed collaborative filtering algorithms)
System : Making good, useful recommendations (effectiveness of algorithm)
Output (Recommendations) :
Presenting recommendations quickly enough but not too quickly (knowing when to say “I can’t recommend”)
Generating trust that system understands user tastes
Providing enough information about each item
Challenges of Recommender System design
Domain differences drive design
Form of sample (song clip vs. product description vs. full text article)
Genres: how fixed and predictable are they?
Frequency of updates (e.g., news & other fast-flowing content)
Commerce vs. taste exploration vs. info-seeking
Some observations & design principles
Trust is crucial
Users think recommender systems have personalities
First impressions are crucial
Does system understand me?
Should I act on its recommendations?
Two different approaches:
Amazon offers affirming experience: familiar items may be correct but not as useful (not new information)
MediaUnbound : less familiar, so more salient and possibly serendipitous, but less likely be acted upon
Source: Sean McNee, John Riedl, Joseph Konstan, CHI Proceedings 2006 “ Making Recommendations Better: An Analytic Model for Human-Recommender Interaction”
Make system logic transparent
Users want to understand why an item was recommended to them
To decide whether to accept recommendation
Identify the input for particular recommendation
How to motivate participation
Easy & engaging process for giving input (MediaUnbound)
Ask at the right moment (Netflix)
Give users control…
Offer filter-like controls for genres/ topics.
Ask how familiar recs should be
Provide detailed info about recommended items
Design principle: Provide clear paths to detailed item information and community feedback such as
Ratings by other users
Sample of item
The unfulfilled promise of Recommender Systems
Some very popular systems (Amazon & Netflix)
Overall, recommender systems lost steam—nowhere near as popular as search.
Data sparseness (unlike search which builds on preexisting data – hyperlinks)
Cold start problem
Gaming the system / spam etc.
Hard to understand and control
Lacked a larger purpose; an end in themselves
Source: Paolo Massal and Bobby Bhattacharjee, Proc. of 2nd Int. Conference on Trust Management, 2004 “ Using Trust in Recommender Systems: an Experimental Analysis”
Recommendations Circa 2006
What’s happened in the interim?
Social networking systems (Friendster, Orkut, LinkedIn, MySpace)
Tagging / folksonomies
Rich interfaces (AJAX / Flash)
People read, write, play, share pics, videos on the web. They live their lives on the web.
Pandora as a textbook example of recommender design principles
Characteristics of Pandora
Rich interface makes experience seamless
Starts giving results with one click
Puts user in control of recommendation
Takes a conversational tone
Not scalable approach
Not social approach: feels like a machine doing thinking for me
Last.fm: a social approach to recommendations
Exploring music at Last.fm
Characteristics of Last.fm
Quick start, friendly interface
Multiple points of entry: charts, tags, users, new items - not just what system recommends for you
Focus on social approach
Listen to other users’ radio stations (Friends, Neighbors, Groups)
Chat on message boards
Highlights contributions to system: your radio station is available to others
Other social recommenders…
What do these systems have in common?
User-generated content: mass participation & social sharing
User- curated content: tags, collections etc.
Harnessing wisdom of crowds
Granular addressability of content
The long tail: making the esoteric more findable
Incorporating social networks
Rich user experience
Not all work: elements of fun and play
Tim O’Reilly, “What is Web 2.0: Design Patterns and Business Models for the Next Generation of Software”
A revolution in RS user experience
User interacts with algorithm to get recommendations
System may use aggregated data about other users (via collaborative filtering algorithms). That data is not directly accessible to all
Centered on completing a finding task or making sales
User interacts with other users, their content and tags to find information & connect with people
Data from other users is exposed and updated in real-time
Succeeds by building a social web, making it more like an ongoing conversation than a transaction
2001 2006 Intelligent Agents Information & Social Hubs
User experiences for finding
User experience with social recommender systems
Move at a slower pace
Get the lay of the land,
Choose paths – what is promising, what sights lie on the way, how well worn.
Easy to change directions, change paths, create your own path
Flickr photostream: soundfromwayout
Design Principle 1: Make system personally useful (before recommendations)
System should serve other useful purpose before it starts personalizing
Portable storage (photos, bookmarks)
Aggregate popular news stories & feeds
Offer vehicle for trendsetters / trendspotters
Provide a discussion forum
Personalize once system has user data
Solves input problem of early RS
Del.icio.us is useful from saving first link
Design Principle 2: Make system participatory
Artistic expression (Flickr, YouTube)
Beyond rating items – contributions of tags, comments, items
Different types of participation
Social software sites don’t require 100% active participation to generate great value.