August 21, 2006
Bartell results
First tried it with the target ground truth being the PPK kernel on uspop37. Did
pretty awfully on the R-precision task, about .16 (a random kernel with the correct diagonal gets about .125).
Using the LFM-collection ground truth and uspop1000, the results were much better. Below is a plot of R-precision vs. Bartlell's 'k' parameter, which is the dimensionality of the semantic space.
Also, the LFM ground truth kernel by itself gets very high R-precision: .567, max hubness 325.
Ah ha... But if I normalize the collections before computing the kernel, then the Bartell results are much better:
Bartell experiment HOWTO
# train the full-cov single gaussian models:
./run-stats.pl -list libraries/uspop_computed_rand1000.list -models uspop1000-fcvg.models -pfile_dir /homes/madadam/projblush/musicsim/uspop/pfiles
# get the collections from LFM by running the scripts in nyab:work/data/lfm/ :
# do-fetch; run-scraper-local; run-filter-lo;
# copy here to 'collections-uspop1000.txt'
In matlab, do_bartell.m