http://research.microsoft.com/pubs/259290/GraphView%20User%2...:

  CREATE TABLE EmployeeNode
  (
    [ColumnRole:”NodeId”] 
    WorkId varchar(32),
    [ColumnRole:”Property”]
    Name varchar(32),
    [ColumnRole:”Edge”, Reference:”EmployeeNode”]
    Colleagues VARBINARY,
    [ColumnRole:”Edge”, Reference:”ClientNode”, Attributes: {credit:”int”, note:”string”}]
    Clients VARBINARY
  )

=> they also appear to replace "many rows each with key and a single referenced value" by "row for each unique key with packed values for that key" (possibly with multiple rows for a key if the packed data doesn't fit in a VARBINARY(MAX))

So, the database structure looks a bit like an inverse index. That could speed up some queries.

I have been playing with something similar in pure T-SQL, using a cross apply over a table-valued function to get from this packed format back to the normalized table, but wasn't impressed with its performance. Using CLR was next on my list, but I haven't come around it because I don't have high hopes that it will perform, but I may try this.