On 2017-05-17 17:08, Zenaan Harkness wrote:
Has anyone done anything like this, and if so, how did you solve it?
(Medium term, the problem begs for a git-like addressed, git-like P2P
addressable/verifiable "content management" solution; e.g. if I have
a random collection of say 10K books, it would be nice to be able to
say something like:
git library init
# add my books:
git add .
git commit
git library sort categories
git library add index1 # create an "indexing" branch
git commit
# add some upstream/p2p libraries for indexing/search/federation:
git library index add upstream:uni.berkely
git library p2p add i2p.tunnel:myfriend.SHA
git library full-index pull myfriend.SHA
git library math-index pull myfriend.SHA
git library book:SHA pull myfriend.SHA
This is the problem of clustering in groups of enormously high
dimension, which is a well studied problem in AI.
Take each substring of six words or less, that does not contain a full
stop internally. The substring may contain a start of sentence marker
at the beginning, and or an end of sentence marker at the end.
Each substring constitutes a vector in a space of enormously high
dimension, the space of all possible strings of moderate length.
Each such vector is randomly but deterministically mapped to a mere
hundred or so dimensions, to a space of moderately high dimension, by
randomly giving each coordinate the value of plus one, minus one, or zero.
Two identical substrings in two separate documents will be mapped to the
same vector in the space of moderate dimension.
Each document is then given a position in the space of moderately high
dimension by summing all the vectors, and then normalizing the resulting
vector to length one.
Thus each document is assigned a position on the hypersphere in the
space of moderately high dimension.
If two different documents contain some identical substrings, will tend
to be closer together.
We then assign a document to its closest group of documents, and the
closest subgroup within that group, and the closest sub sub group.