fixed problem parsing answer file

results with update from head

merged in head results

Split out authored_by_me into separate candidate generator

added features about the user

multithreading fixes; seems to work properly now

multithreaded; still some problems to debug with missing feature vectors

new test with 2000 candidates for the in_cluster sources

fixed glz classifier problems; retest

use bagged boosted trees for source classifier as well

re-running on other machine with clustering speedup

updated README with more information 2

added dot prod features; doesn't seem to make much difference

tried to fix up in_cluster_repo by partially undoing changes

in_cluster_repo fixed up'' (but results are much worse)

screwed up user clusters with too many extra features

new filtering of user coocs

results with classifier changes but no cluster changes yet

coocs and coocs2 improvement

by_watched_authors improvements

first real results with sources restructured

half way through restructure

used bagged boosted stumps

fixed bbdt classifier

touch of results.txt

probabilistic output format then post-processed to get normal results; back to one layer of classifiers

back to bagged boosted trees; still one-level classification

trivial change to get a score

first use of keywords (very quick and dirty)

added repo data up to here

both cooc and cooc2 added

cooccurrences number 2

added cooccurrence suggestions; big difference in range but not much in top

improved readme