At the UIE Web App Summit in Monterey, Rashmi Sinha walked through a series of design strategies for two types of recommender system designs: algorithmic systems prevalent in 2001 and social systems popular in 2006.
- Circa 2001 recommender systems primarily use information about users to predict what may interest them. “If you liked this, you may like…”.
- Circa 2006 recommender systems help people find what may interest them through social or other connections.
Recommender Systems circa 2001
- What movies to watch, books to read, websites to visit, etc.
- Collaborative filtering uses common inputs (like ratings) to correlate likeness between users/objects.
- Trust is crucial: First impression tell users if system understands them.
- Make system logic transparent: explain why an item was recommended.
- Motivate participation: easy & engaging process with a mix of different types of questions and continuous feedback.
- Give users control: over genres or topics; over how familiar or far out recommendations are.
- Detailed information: provide clear paths to detailed item information and reviews, ratings, and samples.
- Overall, most of these types of recommender systems lost popularity to search: sparse data, bad starts, spam, hard to control, etc.
Recommender Systems circa 2006
- People now live their lives on the Web: blogs, wikis, social networking, etc.
- Social recommenders (last.fm, flickr, YouTube, del.icio.us, etc.) have: user-generated content; long tail content; social networking; rich user experience; elements of play.
- In 2001, users interacted with algorithms based on aggregated data. In 2006, users interact with other users making recommender systems more like a conversation than a transaction.
- Make system personally useful: serve a useful purpose (i.e. store content) before you begin personalizing.
- Make system participatory: enable “bite-sized self expression” and go beyond rating systems (tags, comments, implicit creation, remixing, etc.)
- Making participatory process social: real time updates and user profiles/photos make it feel like a conversation.
- Instant gratification: provide personalized recommendations when given input.
- Cultivate user independence: to prevent “mobs” optimize the wisdom of crowds through cognitive diversity, independence, decentralization, and easy aggregation.
- Provide access to long tail: recommend lots of different things and enable fast movement (what’s popular now).
- Expose metadata: expose tags & users and allow people to pivot on these.
- Provide balance between public & private: allow filtering by topic, general likes/dislikes, resetting of profiles.