I'm trying to wrap my head around locality-senstive hashing in the case when you can not enumerate all possible features (e.g. Facebook likes when comparing users). Are there solutions adressing this problem?
The Locality-sensitive hashing algorithms I've seen so far depend on finite vectors of length k where k is the total number of features (e.g. words). In my case, I do not know the total set of features beforehand, still I want to find the n-nearest neighbours for the new item in my database.
Given the targeted size of the database, re-calculating the pairwise similarity for each insert is not feasible.
How can I tackle this problem? Has anyone encountered a similar problem and found a solution?
