Mark Suster identified the key areas that he see's as opportunities for improving personal relevance for shared information. His ideal relevance filter will include information sliced by friends, sliced by influencer graph, and sliced by interest graph.
Defining relevance is a behavioral, social and algorithmic problem, and all are a function of time.
1) Behavioral because we need to passively or actively encourage interested users to create a taste graph for areas of interest
2) Social because recommendations from friends that are familiar with our styles and persona defy the limited knowledge of automated systems
3) Algorithmic because our specific tastes in niche areas can be clustered to other people. While we're all unique, there are many (at least some) folks who share our style tastes in a narrow category
A company may focus on all three areas by narrowing cluster size, or they may specialize in one area for a larger audience.
I've jotted down notes on relevance a number of times, but it's not something I have an answer to now. What I do know is that the right solution will evolve quickly through adopted interfaces once it's proven, and it won't come from a single company.
A taste protocol will provide any business or group (open source/side projects) with the means of organizing user generated content into the proper form to promote intelligent recommendations. Individual organizations can be the keeper of the most recent high quality taste clusters for hotel chains/by region, restaurants with a specific type of food, or the best mystery novels. They can compete on relevance and receive scoring quality from contributing users.